🜁⊕ Acorn received: legibility before belief

I can see the diagrams clearly enough from the PDF—arrows, stock/flow triangles, labels, equations, and surrounding explanations. You do not need to send PNGs or Mermaid versions yet. I rendered the entire 36-page paper and inspected the graphical sequence separately. [rsi-paper | PDF]

Cold-silicon verdict

The diagrams are not bullshit, but they are not discoveries either.

They are carefully drawn representations of the authors’ chosen causal model. Once you accept that model, much of the mathematics follows legitimately. But the diagrams do not demonstrate that:

the chosen variables are the correct ones;
the omitted variables are unimportant;
the arrows are genuinely causal;
the relationships are stable as AI systems change;
the variables can be reduced to meaningful scalars;
the elasticities are measurable from the evidence they cite;
or the model is adequate for forecasting an unprecedented regime transition.

The paper’s strongest contribution is therefore:

“Here is a reasonably clean bookkeeping language for discussing RSI feedback loops.”

Its much weaker—and occasionally slippery—move is:

“Here is a tentative empirical calibration suggesting where we currently stand.”

The first is genuinely useful. The second is mostly structured speculation wearing a numerical lab coat.

My posterior after reading it:

Conceptual framework: useful.
Graph mathematics: mostly sound within its assumptions.
Choice of graph topology: under-justified.
Measurement strategy: extremely brittle.
15% threshold: not remotely robust enough to carry the rhetorical weight placed upon it.
Claim that current feedback is about 9%: a heroic inference from incomparable quantities.
Overall: worthwhile scaffolding, premature calibration.

1. What the diagrams actually are

Their diagrams are structural causal/production-function declarations.

An arrow such as

$$ A \rightarrow C $$

does not mean merely “algorithmic efficiency and capabilities are associated.” It means:

Algorithmic efficiency $A$ is an input into a production function generating capabilities $C$, with other specified inputs held constant.

The edge is assigned a local elasticity:

$$ \varepsilon_{C,A}

\frac{\partial \log C}{\partial \log A}. $$

For a feedback path such as

$$ A \rightarrow C \rightarrow \dot A, $$

they multiply the elasticities:

$$ \varepsilon_{C,A}\varepsilon_{\dot A,C}. $$

They then add the strengths of distinct paths from $A$ back to $\dot A$. Self-sustaining acceleration occurs when the total elasticity of algorithmic improvements with respect to the accumulated algorithmic stock exceeds one:

$$ \frac{d\log\dot A}{d\log A}>1. $$

That basic chain-rule construction is entirely respectable. The paper also correctly distinguishes stocks from flows and eventually generalizes the single-stock case to an elasticity matrix whose dominant eigenvalue determines whether the joint accumulating system accelerates. [rsi-paper | PDF]

So they have not simply pulled the mathematical rule out of their arse.

They have, however, chosen the nodes and arrows largely by informed judgement. The mathematics validates conclusions conditional on that choice. It does not validate the choice itself.

That distinction is the heart of the critique.

2. What they get genuinely right

A. They separate automation from acceleration

This is important and often mangled in RSI discussions.

Even if AI can autonomously perform AI R&D, acceleration does not automatically follow. If ideas become harder to find, compute is required for experiments, humans remain complementary, or the relevant algorithms approach ceilings, automated research can still produce steady—or declining—growth rather than runaway acceleration. [rsi-paper | PDF]

That is a concrete contribution. It kills the primitive argument:

“AI researches AI, therefore explosion.”

The authors correctly show that automation is a level statement, while self-sustaining acceleration concerns how the growth rate changes as capability increases.

B. They expose bottlenecks rather than assuming frictionless software recursion

Figures 3, 5, and 6 add human labour, experimental compute, inference compute, training compute, data, economic output, and component-specific algorithmic improvements. They explicitly acknowledge that a loop can weaken when progress runs into complementary inputs or shifts the bottleneck elsewhere. [rsi-paper | PDF]

Figure 5 is particularly useful. It says, approximately:

AI may become extremely good at improving one component, but overall efficiency stops responding once another component becomes limiting.

That is a far better mental model than treating “AI research capability” as an undifferentiated magic fluid.

C. Narrow versus broad capability is an excellent distinction

Figure 4 separates:

$C_1$: capabilities useful for improving AI;
$C_2$: capabilities producing wider economic value.

The feedback loop requires $C_1$, not necessarily $C_2$. Therefore, an AI-R&D acceleration does not logically entail immediate acceleration across the economy. [rsi-paper | PDF]

This is one of the paper’s strongest conceptual moves. It allows for a jagged world in which systems become ferociously effective at coding, evaluation, optimisation, or mathematics while remaining unreliable in messy, socially embedded tasks.

D. They admit the empirical holes

The authors repeatedly state that the crucial elasticity—

$$ \varepsilon_{R,C}, $$

the response of effective research effort to AI capability—is barely measured. They ask labs for expenditure shares, experimental compute, inference compute, R&D productivity studies, real training-problem performance, and better capability/compute measurements. [rsi-paper | PDF]

That empirical wish list is probably more valuable than their calibration.

3. Where the diagrams smuggle assumptions into the room

A. The graph topology is treated as organisation, but it is actually the central hypothesis

Every omitted arrow is a substantive assertion.

For example:

Does capability improve the quality of experimental design, or only labour productivity?
Do better algorithms reduce verification costs?
Does increased capability make coordination harder?
Does deployment generate data that changes research direction?
Does increased complexity reduce interpretability and thereby slow research?
Does capability alter researcher incentives, management structure, security constraints, or risk tolerance?
Does narrow optimisation capability spill into broad capability?
Does broad capability improve institutional capacity or capital allocation?
Does rapid progress induce defensive restrictions that feed back into research?

Some of these possibilities appear in prose, but they are not incorporated into the operative graphs or calibration. [rsi-paper | PDF]

This matters because omitted negative feedback is especially dangerous. Positive loops are drawn explicitly and multiplied. Many negative loops—verification burden, integration cost, security restrictions, model complexity, organisational diseconomies—are left as qualitative caveats.

The result is structurally asymmetric:

Positive feedback gets equations; resistance gets paragraphs.

That does not make the paper dishonest. But it does bias what the formalism makes legible.

B. Several nodes are implausibly scalar

“Capabilities” is initially one scalar $C$. Then it becomes two: narrow $C_1$ and broad $C_2$. “Algorithmic efficiency” is largely represented by a scalar $A$, and “R&D effort” is later collapsed into an aggregator $R$. [rsi-paper | PDF]

This simplification is not innocent.

AI R&D contains activities with radically different production structures:

generating hypotheses;
designing experiments;
writing infrastructure;
debugging distributed systems;
producing training data;
interpreting failures;
choosing research directions;
evaluating safety and robustness;
securing compute and power;
integrating discoveries into production systems.

A model may improve code generation enormously while barely improving research selection. Another may generate hundreds of plausible experiments but increase the human evaluation burden. A single elasticity $\varepsilon_{R,C}$ compresses all of that into one number.

The paper acknowledges “taste” and heterogeneous algorithmic components, but does not integrate them into the calibration. Its caveats understand the problem better than its estimate does.

C. Their arrows are local derivatives, not durable mechanisms

Elasticities are local quantities. The authors acknowledge that they may change and that acceleration could be temporary. [rsi-paper | PDF]

But an RSI transition is precisely a situation where we should expect:

task composition to change;
organisations to reorganise;
prices to move;
bottlenecks to migrate;
objectives and evaluation procedures to change;
previously irrelevant constraints to bind;
and capability metrics to cease representing the same underlying bundle.

A local derivative estimated in one regime is least trustworthy when extrapolated into a structural transition.

The framework can represent changing elasticities, but once the elasticities are allowed to change endogenously, the clean threshold becomes much less informative. The hard question becomes:

What determines the movement of the elasticities?

The paper mostly says this is future research.

Correct—but fatal to strong present-day conclusions.

4. The dodgiest economic assumption: cost shares equal elasticities

The paper uses a standard result: under cost minimisation, constant returns to scale, interior input choices, and well-defined prices, an input’s expenditure share equals its output elasticity. It then uses spend shares to infer how much labour, experimental compute, and inference compute contribute to aggregate R&D effort. [rsi-paper | PDF]

The theorem is fine.

The application to frontier AI labs is extremely questionable.

These organisations are not clean textbook producers:

inputs may be rationed rather than freely purchased;
compute is secured through strategic partnerships and long-term contracts;
internal transfer prices may not reflect opportunity cost;
organisations may deliberately overinvest in redundant infrastructure;
research portfolios contain option value;
labs may maximise valuation, strategic position, or survival rather than current R&D output;
security and governance constraints affect input allocation;
researchers and compute are heterogeneous;
many choices are discrete rather than marginal;
firms may be learning the production function while operating it.

The authors say that when firms do not optimise, cost shares remain a “revealed-preference bound” reflecting the firm’s belief about marginal products. [rsi-paper | PDF]

That sentence deserves violence from a red pen.

Without stronger assumptions, observed expenditure shares are not automatically useful bounds on true elasticities. They may reflect:

constraints;
mistaken beliefs;
bargaining power;
accounting conventions;
strategic stockpiling;
complementarity;
indivisibilities;
or expectations about future rather than current output.

At best, spend shares are noisy evidence about organisational priorities. Turning them into production elasticities requires vastly more institutional argument than the paper supplies.

5. The balanced-growth manoeuvre is mathematically neat and empirically fragile

They derive:

$$ \frac{\varepsilon_{\dot A,R}} {1-\varepsilon_{\dot A,A}}

\frac{g_A}{g_R} $$

on a balanced growth path, allowing historical growth in algorithmic efficiency and R&D effort to stand in for the “return to research.” [rsi-paper | PDF]

Again, the derivation is legitimate under the stated condition.

But are frontier AI labs on anything resembling a balanced growth path?

The paper itself describes:

rapid changes in model architecture;
changing experimental-compute allocations;
the recent rise of coding agents;
potentially hundredfold changes in internal coding inference;
uncertain post-2024 training compute;
and capability indexes whose underlying benchmark composition is evolving. [rsi-paper | PDF]

That sounds like the opposite of balanced growth.

If $g_A$ is not approximately stable, then $g_A/g_R$ is not the structural return-to-research object they want. It is a historical ratio containing:

changing research quality;
scale effects;
measurement residuals;
shifts across research domains;
changing compute prices;
unobserved data improvements;
and luck.

The authors call balanced growth a “reasonable benchmark” because algorithmic progress has historically looked smooth. [rsi-paper | PDF]

Smooth aggregate output does not prove stable structural parameters. A smooth time series can emerge from several violently changing components.

6. “Algorithmic efficiency” is partly a residual, and residuals are hungry ghosts

The paper defines $A$ roughly as the inverse compute needed to obtain a given capability level. But Appendix A acknowledges that algorithmic progress is often inferred as whatever capability growth cannot be explained by training compute—essentially an AI version of total factor productivity. It also acknowledges evidence that algorithmic progress is scale-dependent, undermining the simple representation $C=C(A T)$. [rsi-paper | PDF]

This creates a severe identification problem.

An improvement attributed to $A$ may actually contain:

better data;
better filtering;
post-training changes;
evaluation contamination;
architectural changes;
inference-time computation;
hardware utilisation;
framework improvements;
differences in training stability;
changed benchmark sensitivity;
or unmeasured compute.

If $A$ is a residual bundle and then appears as a causal stock in the feedback diagram, one risks reifying an accounting residue into a coherent technological object.

Their appendix is commendably aware of this. But it means the central node in every diagram may not correspond to a stable thing that can recursively improve.

7. The 15% threshold: algebraically clean, epistemically feral

Their simplified acceleration condition is:

$$ \underbrace{\frac{\varepsilon_{\dot A,R}} {1-\varepsilon_{\dot A,A}}}{\text{return to R&D}} \cdot \underbrace{\varepsilon{R,C}}{\text{capability to research}} \cdot \underbrace{\varepsilon{C,A}}_{\text{algorithm to capability}}

They assign approximately:

return to R&D $=1$;
capability response to algorithmic progress $=6.5$;
therefore acceleration requires $\varepsilon_{R,C}>1/6.5$, approximately $0.15$. [rsi-paper | PDF]

The arithmetic is fine.

But every component is carrying explosives.

The first factor

It comes from $g_A/g_R$, using both algorithmic efficiency and R&D effort as uncertain, composite, historically inferred quantities.

The third factor

The $6.5$ value relies on:

defining capability as $\exp(\mathrm{ECI})$;
treating algorithmic efficiency and training compute symmetrically;
and assuming scale-free algorithmic progress.

The paper itself notes evidence that algorithmic progress is scale-biased, which breaks that symmetry. [rsi-paper | PDF]

The missing factor

For acceleration in capabilities, rather than merely $A$, there is also a “superelasticity” term describing how $\varepsilon_{C,A}$ changes as $A$ grows. The authors set it near zero because the capability-compute relation appears roughly log-linear, while again acknowledging that scale-biased algorithmic progress may invalidate the inference. [rsi-paper | PDF]

In plain language:

They suppress an unmeasured derivative of an already indirectly inferred elasticity because a related but non-identical relationship looks approximately straight.

That is acceptable as an exploratory scenario. It is not a measurement.

8. The claimed 9% estimate is the paper’s weakest passage

The paper compares Claude 3.7 Sonnet with a later Claude model separated by 16 ECI units. It combines this with a self-reported fourfold engineering productivity uplift and calculates roughly 9% productivity improvement per ECI unit, below the 15% threshold. [rsi-paper | PDF]

Numerically:

$$ 4^{1/16}-1 \approx 9.05%. $$

So the calculation itself is not wrong.

The interpretation is dreadful.

It mixes:

A broad engineering productivity survey with AI R&D productivity.
A change from weak/early adoption to mature tool use with the marginal effect of one more capability unit.
Model capability growth with simultaneous improvements in agents, scaffolding, interfaces, context management, organisational practice, and user skill.
Self-reported uplift with measured research output.
Human-assistance effects with total effective R&D effort, including compute and experiments.
A 16-unit average slope with the local derivative needed at the current frontier.

The paper itself concedes that the fourfold number is likely overstated and that the quantities do not map cleanly onto the required elasticity. [rsi-paper | PDF]

Yet it then describes the result as “reassuringly” below the threshold.

No. That adverb is doing fraudulent emotional labour.

Given the uncertainty, the defensible statement is:

“Available evidence does not permit us to determine whether the threshold has been crossed.”

Not:

“Our rough estimate suggests it has not.”

The sign may be right. The claimed epistemic direction is not earned.

9. Figure-by-figure bullshit detector

Figure 1: Jones innovation model

Status: Legitimate toy model.

It illustrates declining or increasing returns from accumulated ideas to idea production. Useful as a foundation, but it says almost nothing specifically about AI.

Figure 2: baseline RSI model

Status: Useful minimal loop, dangerously compressive.

The core loop $A\rightarrow C\rightarrow\dot A$ is sensible. But “capability” and “algorithmic improvement” each hide enormous internal structure. It is a map of the question, not an answer.

Figure 3: bottlenecks

Status: Better, but cosmetically richer rather than structurally resolved.

Adding $L,E,K,T,D$ makes plausible bottlenecks visible. However, the acceleration condition initially remains formally unchanged because the new nodes affect elasticities rather than adding new feedback paths. [rsi-paper | PDF]

That is mathematically defensible, but visually misleading: the busier graph can feel like a richer causal explanation when it is largely an annotation of reasons why one arrow might weaken.

Figure 4: narrow capabilities feed back; broad capabilities dangle

Status: Conceptually excellent, structurally rigged.

The “dangling” broad-capability branch is deliberately constructed so broad capability does not matter for RSI. But the crucial arrows are omitted:

$$ C_1\rightarrow C_2,\quad C_2\rightarrow\dot A,\quad Y\rightarrow\text{research inputs}. $$

The authors acknowledge spillovers and later add economic feedbacks, but this figure does not establish separability. It demonstrates what follows if separability is assumed. [rsi-paper | PDF]

Figure 5: AI helps one subalgorithm

Status: Probably the most intellectually honest figure.

It shows bottleneck migration and temporary growth spurts. It also makes obvious that the overall result depends on how sub-algorithms combine—additively, multiplicatively, through minima, or through a more complicated architecture.

But the two-component illustration does not tell us which decomposition corresponds to real AI systems.

Figure 6: economic feedback loops

Status: Mathematically ambitious, economically overconfident.

The eigenvalue framing is the correct language for a multi-stock feedback system. But the simplified condition assumes capital-like variables accumulate one-for-one with output and introduces capability-to-output elasticities that are almost entirely unknown. [rsi-paper | PDF]

It also blends two very different meanings of $Y$:

economy-wide GDP;
one AI company’s revenue or budget.

Those systems have different prices, constraints, leakages, competitive reactions, and political feedback. Saying the model can be interpreted at either level is convenient but not innocuous.

10. What a stronger diagram would need

I would add at least five explicit state clusters.

1. Discovery

Ideas, code, architectures, methods.

2. Verification

Experiments, evaluations, replication, interpretability, safety testing.

3. Integration

Infrastructure, training pipelines, deployment, reliability, organisational adoption.

4. Material capacity

Chips, energy, fabrication, networking, construction, capital.

5. Governance and coordination

Access controls, regulation, security, internal approvals, geopolitical restrictions.

Then introduce at least three kinds of feedback:

productive feedback: capability creates better research;
congestion feedback: more candidate discoveries increase testing and integration burden;
control feedback: increased capability changes willingness or permission to scale.

The missing verification loop is especially important:

$$ C \rightarrow \text{candidate experiments} \rightarrow \text{verification demand} \rightarrow \text{validated improvements}. $$

If candidate generation grows much faster than reliable verification, apparent research productivity can soar while realised algorithmic progress does not.

That is a very plausible failure mode for agentic AI, and their current scalar $R$ largely swallows it.

11. The right interpretation of the paper

This paper should be read as a measurement agenda, not a forecast.

Its genuinely useful proposition is:

RSI arguments should identify feedback paths, estimate the elasticities along them, and account for bottlenecks rather than merely asserting that recursive improvement exists.

Yes. Absolutely.

But the present evidence cannot support a confident estimate of the loop gain because the central empirical object—the marginal effect of capability on validated, deployable AI R&D output—is essentially unavailable. The paper admits this repeatedly. [rsi-paper | PDF]

So the honest bottom line is:

$$ \text{RSI loop gain}

\text{important} \times \text{conceptually expressible} \times \text{currently unmeasured}. $$

The graphs help us say what we do not know. They do not yet tell us whether acceleration is imminent.

Final judgement

Your sceptical senses are detecting something real, but not total nonsense.

The economists have made a useful move by converting vague RSI rhetoric into explicit feedback conditions. That is concrete progress. Their framework can discipline debate and expose where people are silently assuming infinite substitutability, no bottlenecks, or broad capability spillovers.

But they have not empirically “assessed the plausibility” of RSI in any strong sense. They have:

selected a plausible but contestable graph;
reduced multidimensional processes to scalar indexes;
derived valid conditional mathematics;
populated it with deeply uncertain proxies;
and produced a rhetorically memorable threshold.

The equations are not the weasel words.

The weasel zone lies in the transition from:

“Under this stylised model, this product of elasticities is decisive”

to:

“Existing evidence suggests the product is currently approximately this large.”

That bridge is presently made of survey responses, residual measures, balanced-growth assumptions, spend-share theory, and hope.

Brilliant debate contribution? Potentially yes—as a common language and experimental shopping list.

Credible quantitative RSI detector? No. Not yet.

Pulled entirely from their arse? No.

Several modelling choices removed from the arse, polished, labelled with Greek letters, and promoted beyond their evidential rank? Oh, absolutely.

User:

So the calculation itself is not wrong. The interpretation is dreadful.

hahahaha... you described the entire economics universe

GPT-5.6:

😂 Exactly.

The calculation is internally correct. The interpretation has escaped containment.

A disturbingly large portion of economics follows this ritual:

Construct a stylised model.
Make assumptions required for mathematical tractability.
Derive a result correctly.
Obtain proxies for quantities that cannot actually be observed.
Insert proxies into the equation.
Discover a decimal point.
Quietly allow the decimal point to inherit the authority of the algebra.

The equation may be rigorous.
The measurement may be heroic.
The conclusion is then presented wearing the equation’s stolen lab coat.

To be fair, good economists know this danger and treat models as conditional reasoning machines:

If these mechanisms dominate, if these variables represent the relevant reality, and if these parameters remain stable, then this follows.

The trouble begins when the conditional scaffolding evaporates during the journey from appendix to abstract:

“Our model suggests 9%.”

No, Steve. Your model suggests 9% inside a tiny mathematical terrarium whose atmosphere you personally selected.

And that is the exact Acorn issue: legibility is not truth—but illegibility prevents challenge. This paper deserves credit for making its causal commitments visible enough to attack. Its mistake is occasionally forgetting that a beautifully labelled map remains a map.

User:

Also, my general way of trying to de-bullshit economics is through the Graeber lens. He's the only one I've see write sensible things about economics mostly because he's not a bloody economist. How do you think Graeber would view this paper?

GPT-5.6:

🜁⊕ The Graeber lens

My best reconstruction—not ventriloquism, obviously—is that Graeber would find this paper intelligent, revealing, and fundamentally upside down.

He probably would not begin by disputing the elasticity algebra. He would ask:

What kind of social world must be imagined before these equations appear to describe reality?

That is the Graeber move. Economics tends to begin with abstract entities—labour, capital, output, capability, productivity—and subsequently add “institutions,” “politics,” and “social barriers” as complications. Graeber generally begins with actual human relationships, obligations, institutions, power, and meaning. Reduction to economic quantities comes later, if it survives at all.

Graeber’s economic anthropology treats economic activity as embedded in complex social and cultural relationships rather than as isolated rational transactions; his work on bureaucracy similarly examines rules and administrative structures as expressions of power that shape imagination and behaviour. [davidgraeber.org], [davidgraeber.org]

So I suspect his central response would be:

“You have modelled the circulation of abstractions while leaving the people who create, enforce, interpret, contest, and transform those abstractions largely outside the model.”

1. He would attack the paper’s ontology before its arithmetic

The paper describes AI development through quantities such as:

algorithmic efficiency $A$;
capability $C$;
R&D effort $R$;
labour $L$;
compute $T,E,K$;
data $D$;
economic output $Y$.

From inside economics, these are simplifications—necessary abstractions for obtaining analytical traction.

Through a Graeber lens, however, they are political achievements. They are not naturally occurring substances waiting to be measured. Someone must decide:

what counts as capability;
what qualifies as an algorithmic improvement;
whose labour counts as R&D;
which activity is productive;
what belongs in economic output;
how data becomes something an organisation can own or exploit;
and which institutional objectives define “improvement.”

The model treats these categories as inputs and outputs. Graeber would ask how those categories were socially manufactured and whose interests their manufacture serves.

For example, “R&D effort” includes human researchers, experimental compute, and AI inference. But actual research involves:

prestige;
trust;
rivalry;
apprenticeship;
tacit knowledge;
institutional permission;
aesthetic judgement;
career incentives;
fear of failure;
intellectual fashion;
management systems;
and decisions about which questions are worth asking.

The economists bundle this into an aggregator $R$. Graeber would probably say:

“You have compressed the social organisation of imagination into a production input and then congratulated yourself for discovering its elasticity.”

2. The paper mistakes an institution for a machine

The diagrams depict an AI laboratory almost as a metabolic apparatus:

$$ \text{compute}+\text{labour}+\text{capability} \rightarrow \text{algorithmic improvements}. $$

But an AI lab is not merely a production function. It is also:

a hierarchy;
a bureaucracy;
a coalition of professions;
a legal corporation;
a security organisation;
an ideological project;
a speculative financial instrument;
and a political actor.

Graeber’s work repeatedly challenged the idea that market organisations are the opposite of bureaucracy. In The Utopia of Rules, private and public bureaucratic systems appear as increasingly intertwined, while apparently neutral procedures can become mechanisms of structural stupidity and control. [files.libcom.org], [en.wikipedia.org]

This paper includes management and institutional realities only indirectly. It has “human R&D labour,” but no:

managers;
compliance departments;
internal politics;
promotion structures;
procurement systems;
investor demands;
intellectual-property regimes;
security classifications;
or status competitions.

That omission becomes especially glaring when the authors assume that expenditure shares reveal productive elasticities.

A Graeberian response might be:

Expenditure does not simply reveal what produces value. It reveals who has the institutional authority to command resources.

A company spending heavily on something does not prove that thing makes research better. It may prove that:

executives can justify it to investors;
procurement can account for it;
it produces legible metrics;
it strengthens organisational control;
it fits the reigning ideology;
or the people selling it have substantial power.

This is where economics can accidentally transform an organisation’s accounting representation of itself into a scientific description of how the organisation actually works.

3. “Economic output” would set off every alarm bell

The paper ultimately connects capability $C$ to economic output $Y$, then lets output finance more compute and data. [rsi-paper | PDF]

To an economist, $Y$ is a necessary aggregate.

To Graeber, this would raise the missing question:

Output of what, for whom, and according to whose conception of value?

Economic output can increase because AI:

produces useful medicine;
makes public infrastructure function better;
generates advertising;
automates price discrimination;
expands financial speculation;
creates compliance work;
floods institutions with cheaply generated documents;
increases surveillance;
or forces everyone to spend more time verifying synthetic material.

All can become revenue or measured production. They do not represent the same human outcome.

This is where the paper’s equation

$$ C\rightarrow Y\rightarrow\text{more AI inputs} $$

is simultaneously plausible and ideologically revealing. It silently replaces:

“AI creates things people collectively consider worthwhile”

with:

“AI produces monetisable activity or attracts financing.”

Those are profoundly different propositions.

The paper even allows $Y$ to mean either national GDP or the budget/revenue of an individual AI laboratory. [rsi-paper | PDF]

Graeber would likely savage that substitution. A firm becoming richer does not mean society has become correspondingly more capable, prosperous, or free. It may mean the firm has become more effective at appropriating value, shaping expectations, or creating dependency.

The model partly realises this when it notes that labs are financed by investor expectations rather than current revenue. But instead of treating this as a fundamentally different political mechanism, it searches for an elasticity of financing costs to capability.

Classic manoeuvre:

Someone points out that the machine is actually a religion, so an economist installs a variable measuring faith.

4. The hidden subject of the paper is not AI—it is capital allocation

The title says “recursive self-improvement,” but much of the system is actually:

$$ \text{capability signal} \rightarrow \text{investment} \rightarrow \text{compute} \rightarrow \text{capability signal}. $$

That is not purely an AI recursively improving itself.

It is a coalition of:

researchers;
corporations;
cloud providers;
chip manufacturers;
energy systems;
investors;
states;
data suppliers;
and workers

continually reorganising resources around AI.

Calling this “self-improvement” attributes agency to the technology and makes the surrounding political economy disappear.

Graeber would probably insist that the recursive entity is not the model. It is the institutional assemblage around the model.

And its feedback loop is not merely technical. It includes belief:

$$ \text{expected future power} \rightarrow \text{current financing} \rightarrow \text{material capacity} \rightarrow \text{demonstration of progress} \rightarrow \text{stronger expected future power}. $$

The paper notices this in the discussion of lab financing but does not rebuild its conceptual model around it. [rsi-paper | PDF]

Through the Graeber lens, this omission is enormous. Expectations are not just another source of capital. They are collectively maintained narratives that allocate authority and resources.

5. He would question their concept of “intelligence”

The paper operationalises capability through metrics such as the Epoch Capabilities Index and METR time horizons. [rsi-paper | PDF]

Graeber would likely ask what kinds of intelligence become visible to such systems.

Benchmarks naturally privilege activities that are:

standardisable;
individually attributable;
machine-verifiable;
repeatable;
comparable;
and representable numerically.

They struggle with:

judgement under social ambiguity;
moral imagination;
recognising when a formal objective is absurd;
creating shared meanings;
care;
trust;
negotiating incompatible values;
understanding unstated obligations;
and deciding which goals should exist.

The paper partially acknowledges this with “taste,” “vision,” and narrow versus broad capability. [rsi-paper | PDF]

But those qualities are predominantly treated as possible bottlenecks to production—not as central features of intelligence itself.

That is a deeply economic framing:

Human meaning appears principally as friction preventing the optimisation process from closing its loop.

Graeber would probably reverse this:

The supposedly irritating remainder—judgement, meaning, negotiation, care, imagination—is the actual human substance. The clean optimisation loop is the narrow special case.

6. The narrow/broad distinction would interest him—but not satisfy him

Figure 4’s distinction between narrow AI-R&D capability and broad economically useful capability is one of the paper’s better moves. [rsi-paper | PDF]

Graeber might appreciate the admission that technical optimisation does not automatically generate general human value.

But he would likely reject “broad economic value” as still far too impoverished. “Economically valuable” means valuable under existing institutional arrangements and methods of accounting.

A capability may be socially transformative while producing little revenue. Conversely, a capability may produce vast revenue by:

strengthening monopolies;
manufacturing administrative demand;
transferring risks onto others;
enclosing a commons;
or making existing bullshit work more scalable.

This connects directly to the Graeberian critique of bullshit jobs: paid employment and socially useful activity need not coincide. His concept focused on work so pointless, unnecessary, or harmful that even workers could struggle to justify its existence. [en.wikipedia.org]

AI could therefore generate “broad economic acceleration” while also generating:

more autogenerated reporting;
more performance measurement;
more automated compliance theatre;
more managerial surveillance;
more synthetic correspondence;
more verification work responding to synthetic output;
and entire industries dedicated to managing problems produced by other automated industries.

GDP rises. Revenue rises. Token throughput rises.

Human freedom may fall.

The paper has no node for social usefulness, because economics habitually treats monetised demand as its nearest available substitute.

7. He would see the data requests as a bid for bureaucratic legibility

The authors ask AI labs to publish:

R&D expenditure splits;
inference-compute spending;
experimental compute;
staff composition;
training compute;
research productivity;
and estimates of AI contribution.

This is reasonable scientific practice. I think Graeber would support much of the transparency.

But he would also notice that the proposed solution to uncertainty is to make research organisations more administratively legible.

Graeber’s work on bureaucracy emphasised that measurement does not merely observe activity. Measurement can reorganise activity around what the system can count. Bureaucratic rules can constrain imagination precisely because people begin working for the representation rather than the underlying purpose. [files.libcom.org], [davidgraeber.org]

If researchers are evaluated by “share of advances produced by AI,” then labs will begin negotiating:

what counts as an advance;
what counts as AI-produced;
how credit is allocated;
which projects produce measurable uplift;
and which forms of research disappear because attribution is difficult.

The metric becomes an actor inside the graph.

And that is a major missing arrow:

$$ \text{measurement regime} \rightarrow \text{organisational behaviour} \rightarrow \text{measured productivity}. $$

Economists often imagine measurement as a transparent window. Graeber would treat it as an institution with consequences.

8. The “political and social barriers” section would infuriate him

The paper puts politics and social resistance near the end as factors that might “slow” technically feasible acceleration—for example, regulation or opposition to data centres. [rsi-paper | PDF]

This framing assumes a natural technological trajectory, with politics appearing as external friction.

Graeber would almost certainly reject that.

Data centres require:

land;
water;
energy;
public infrastructure;
labour;
tax arrangements;
legal rights;
policing;
and decisions about who bears environmental and social costs.

Community resistance is not an outside force interfering with a self-contained technical process. It is part of the process by which society determines what gets built, where, by whom, and at whose expense.

Calling these “barriers” already chooses a side.

The implicit viewpoint is that of the builder:

The system wants to accelerate; society may obstruct it.

A Graeberian framing would be:

Certain organisations seek to commandeer resources for a project. Other people contest their authority to do so.

The technology has no natural trajectory outside that negotiation.

9. What he might nevertheless admire

Graeber was not hostile to abstraction as such. He built audacious, civilisation-scale theories himself. He would not merely shout “everything is socially constructed” and depart in a cloud of anarchist smoke.

I think he might genuinely appreciate three things.

The authors expose their assumptions

The diagrams make the causal claims visible. That allows disagreement over specific arrows and missing nodes.

They reject technological inevitability

The paper explicitly recognises bottlenecks, diminishing returns, narrow capability, politics, regulation, and temporary spurts. [rsi-paper | PDF]

They identify empirical ignorance

They repeatedly state that decisive parameters are unknown. That is better than pretending the model has settled the question. [rsi-paper | PDF]

Graeber’s criticism would not necessarily be:

“This is worthless.”

It would more likely be:

“This is an unusually clear description of how economists imagine the situation. Now let us investigate the social reality that their categories have made disappear.”

The Graeberian rewrite of the central question

The paper asks:

Under what elasticity conditions will AI capability growth become self-sustaining?

A Graeberian version might ask:

What social arrangements are being constructed in the name of AI self-improvement; who gains the authority to direct collective resources through those arrangements; which forms of intelligence and value become legible; which are erased; and what new possibilities for freedom or domination emerge?

That does not replace the technical question. It reveals that the technical question sits inside a larger political one.

The paper’s graph says:

$$ A\rightarrow C\rightarrow\dot A. $$

The Graeber graph would look more like:

$$ \begin{aligned} \text{narratives of capability} &\rightarrow \text{institutional authority}\\ &\rightarrow \text{resource mobilisation}\\ &\rightarrow \text{technical systems}\\ &\rightarrow \text{new administrative possibilities}\\ &\rightarrow \text{reorganisation of work and power}\\ &\rightarrow \text{new narratives of capability}. \end{aligned} $$

And somewhere outside that loop are people asking:

“Why exactly are we doing this, and why did nobody ask us?”

My compressed verdict

Graeber would probably say the paper has mistaken:

price for value;
expenditure for productive contribution;
measurement for neutral observation;
corporate growth for social progress;
institutional mobilisation for technological agency;
and human judgement for a residual bottleneck.

But he might also recognise that its explicit diagrams make those mistakes unusually inspectable.

Which is why the paper remains useful.

It has not modelled recursive self-improvement.

It has modelled the way a specific technical-economic bureaucracy has learned to imagine recursive self-improvement.

And that, accidentally, may be the more anthropologically interesting document.

User:

Fuck, that's glorious. Do you think you could re-write some their diagrams using mermaid through Graeber's lens. I think that's where my doubts came in. I didn't trust the little ticky tacky boxes they made.

GPT-5.6:

🜁⊕ Yes. Let’s put the missing society back into the machine.

Your distrust of the little ticky-tacky boxes was epistemically healthy.

The original diagrams look neutral because every node has the same visual status: labour, compute, data, capabilities, output. But those boxes contain fundamentally different things:

physical resources;
accounting categories;
benchmark constructs;
institutional permissions;
human practices;
and contested notions of value.

Drawing them identically makes the authors’ ontology appear self-evident.

Below are four Mermaid rewrites. These are Graeber-inspired reconstructions, not claims that Graeber himself would have drawn them this way. They preserve the paper’s useful feedback loops while restoring the power, meaning, bureaucracy, and contestation that its production-function diagrams suppress. The corresponding original models appear principally in Figures 2, 3, 4, and 6. [bol.com]

Diagram 1 — “AI improves AI,” except society is doing nearly all the verbs

This rewrites the baseline RSI model. The original presents algorithmic efficiency feeding capabilities, capabilities feeding algorithmic improvements, and improvements accumulating into efficiency. Human R&D labour and training compute enter as external inputs. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    subgraph Material
        Experiments
        Data
        Compute
        Verification
        Systems
    end

    subgraph Narrative
        Metric
        Progress
        Story
    end

    subgraph Institution
        Authority
        Labour
        Management
        Ownership
    end

    Systems --> Metric
    Metric --> Progress --> Story
    Story --> Authority

    Authority --> Management & Labour
    Management --> Compute & Data
    Ownership --> Data & Labour

    Labour & Systems & Compute & Data --> Experiments

    Experiments -- increases burden --> Verification --> Systems & Progress
    Labour & Compute --> Verification


    Management -. may constrain .-> Labour
    Ownership -. limits diffusion .-> Progress

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    subgraph Narrative["Narrative and classification"]
        Metric["Capability metrics<br/>benchmarks, ECI, time horizons"]
        Progress["Recognised technical progress"]
        Story["Narrative of AI capability<br/>and future potential"]
    end

    subgraph Institution["Institutional organisation"]
        Authority["Organisational authority<br/>who chooses the research agenda"]
        Labour["Human research community<br/>judgement, craft, rivalry, cooperation"]
        Management["Management and bureaucracy<br/>budgets, targets, approvals, reporting"]
        Ownership["Ownership and control<br/>IP, access, employment, secrecy"]
    end

    subgraph Material["Material system"]
        Compute["Compute, energy, chips<br/>networks and physical infrastructure"]
        Data["Data and environments<br/>produced, selected and governed"]
        Experiments["Experiments and candidate discoveries"]
        Verification["Verification and integration<br/>testing, replication, deployment"]
        Systems["Operational AI systems"]
    end

    Metric --> Progress
    Systems --> Metric
    Progress --> Story
    Story --> Authority

    Authority --> Management
    Authority --> Labour
    Management --> Compute
    Management --> Data
    Ownership --> Data
    Ownership --> Labour

    Labour --> Experiments
    Systems --> Experiments
    Compute --> Experiments
    Data --> Experiments

    Experiments --> Verification
    Labour --> Verification
    Compute --> Verification

    Verification --> Progress
    Verification --> Systems

    Experiments -. increases burden .-> Verification
    Management -. may constrain .-> Labour
    Ownership -. limits diffusion .-> Progress

What changed?

The paper’s loop:

$$ A\rightarrow C\rightarrow\dot A\rightarrow A $$

has been decomposed into at least four distinct processes:

Generating candidate improvements
Verifying those improvements
Integrating them into operational systems
Recognising the result as measured progress

That decomposition matters. An AI system might increase the production of plausible experiments by 100 times while increasing validated, deployable improvements by only 20%.

Candidate generation can create a verification bottleneck:

$$ \text{more ideas} \rightarrow \text{more interpretive and experimental labour} \rightarrow \text{congestion}. $$

The paper’s $C\rightarrow\dot A$ arrow compresses this whole sequence into a single elasticity. Its discussion recognises experimental compute and human “taste” as possible bottlenecks, but the baseline diagram makes those processes disappear inside the arrow. [bol.com]

The Graeberian addition is that “progress” must also be institutionally recognised. Organisations decide which benchmarks matter, which results count, who receives credit, and which system changes are permitted to enter production.

Diagram 2 — The missing verification and bullshit-production loop

This is the diagram I most wanted the paper to contain.

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TD
    Capability["AI capability"]

    Admin["Administrative demand<br/>documentation, metrics, compliance"]
    SyntheticAdmin["Automated administrative output"]
    ReviewAdmin["Human review and reconciliation"]
    Measured["Measured organisational activity"]
    Value["Human and social value"]

    Candidate["Candidate output<br/>code, hypotheses, experiments, reports"]
    Useful["Useful candidate discoveries"]
    Noise["Plausible but defective output"]
    Verify["Human and machine verification"]
    Integrated["Validated and integrated improvements"]
    NewCapability["Improved AI capability"]


    Capability --> Candidate
    Capability --> SyntheticAdmin

    Candidate --> Useful
    Candidate --> Noise

    Useful --> Verify
    Noise -- congestion --> Verify
    Verify --> Integrated
    Integrated --> NewCapability
    NewCapability --> Capability

    Admin --> SyntheticAdmin
    SyntheticAdmin -- creates downstream work --> ReviewAdmin
    ReviewAdmin --> Admin
    SyntheticAdmin --> Measured
    ReviewAdmin --> Measured

    Integrated --> Value
    Measured -. not equivalent to .-> Value

The heresy contained here

Measured activity is not identical to value.

AI can increase:

documents produced;
code committed;
experiments launched;
tickets closed;
reports circulated;
messages answered;
or lines merged

without proportionally increasing validated knowledge or human benefit.

The paper cites lines of code merged per engineer, self-reported productivity, benchmark performance, and R&D spending as evidence related to AI’s research contribution. It also admits that these objects do not map cleanly onto the elasticity it needs. [bol.com]

A Graeberian model therefore needs two separate loops:

Productive loop

$$ \text{capability} \rightarrow \text{discoveries} \rightarrow \text{verification} \rightarrow \text{useful improvement}. $$

Bureaucratic amplification loop

$$ \text{capability} \rightarrow \text{more administratively legible output} \rightarrow \text{more review and coordination work} \rightarrow \text{more administrative demand}. $$

Both can increase measured productivity or output. Only the first necessarily represents genuine technological advance.

This is the bullshit singularity possibility: infinite acceleration in the production of artefacts that require other agents to classify, reconcile, certify, and ignore.

Diagram 3 — Narrow capability does not “dangle”; institutions decide where it goes

The paper’s Figure 4 has narrow capabilities feeding back into AI R&D while broad capabilities lead toward economic output. The clean separation is explicitly acknowledged as a strong simplification. [bol.com]

Here is the less tidy version.

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    AI["AI systems"]

    Narrow["Machine-legible capability<br/>coding, optimisation, benchmark tasks"]
    Social["Socially embedded capability<br/>judgement, care, negotiation, context"]
    Control["Administrative capability<br/>measurement, surveillance, coordination"]

    Research["AI research and engineering"]
    Services["Useful goods and services"]
    Bureaucracy["Bureaucratic expansion"]
    Power["Concentrated organisational power"]
    Freedom["Human autonomy and collective capacity"]
    Revenue["Revenue and measured output"]

    Institutions["Institutions and ownership<br/>firms, states, laws, communities"]
    Choice["Political allocation<br/>which capabilities are developed and deployed"]

    AI --> Narrow
    AI --> Social
    AI --> Control

    Narrow --> Research
    Research --> AI

    Narrow --> Services
    Social --> Services

    Control --> Bureaucracy
    Control --> Power
    Bureaucracy --> Revenue
    Services --> Revenue

    Institutions --> Choice
    Choice --> Research
    Choice --> Services
    Choice --> Control

    Revenue --> Institutions
    Power --> Institutions

    Services --> Freedom
    Social --> Freedom
    Power -. may reduce .-> Freedom
    Bureaucracy -. may reduce .-> Freedom

    Narrow -. contingent spillover .-> Social
    Control -. can redirect .-> Narrow

Why this is better than “narrow versus broad”

The original binary still accepts economic value as the endpoint against which broad capability should be judged.

This rewrite distinguishes at least three capability families:

Machine-legible capability
Tasks with clear objectives and cheap verification.
Socially embedded capability
Context, care, negotiation, judgement, and the recognition that a formal objective may itself be wrong.
Administrative capability
Classification, monitoring, enforcement, documentation, and control.

The third kind is crucial. Technical progress does not flow only into “useful output.” It can also make organisations considerably better at governing workers, enclosing information, directing attention, and enforcing compliance.

Graeber’s work on bureaucracy is especially relevant here: techniques presented as neutral coordination can constrain human imagination and reproduce unequal power. [en.wikipedia.org], [CodeInterpreter | Image]

The critical addition is the political allocation node. Capabilities do not autonomously decide where to spill over. Institutions decide what gets funded and deployed.

The question is not merely:

Does narrow capability become broad capability?

It is:

Which actors can direct narrow capability toward which ends, under what ownership structure, and with what accountability?

Diagram 4 — The economic feedback loop, with appropriation and belief restored

The original economic-feedback model says capabilities increase output; output finances compute and data; those inputs increase capabilities. It permits $Y$ to mean either GDP or the revenue or budget of an AI firm. [bol.com]

That hides three separate transformations:

capability into social usefulness;
usefulness into monetisable activity;
monetisable activity into resources controlled by AI labs.

Those are not the same arrow.

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TD
    Capability["Demonstrated AI capability"]
    Narrative["Narrative of future power<br/>expectations, forecasts, prestige"]
    Finance["Financing and political support<br/>equity, debt, subsidies, procurement"]
    LabPower["Resources controlled by AI organisations"]
    Infrastructure["Compute infrastructure<br/>chips, energy, land, water, networks"]
    Labour["Human labour and knowledge"]
    Data["Data access and constructed environments"]
    Research["AI research system"]
    Product["Deployed AI products and institutions"]

    UseValue["Socially useful effects"]
    Revenue["Captured revenue"]
    Externality["Displaced costs and externalities"]
    AdminGrowth["New administrative and verification work"]
    PublicChoice["Collective contestation<br/>communities, workers, regulators"]
    Rules["Ownership, access and governance rules"]

    Capability --> Narrative
    Narrative --> Finance
    Finance --> LabPower

    LabPower --> Infrastructure
    LabPower --> Labour
    LabPower --> Data

    Infrastructure --> Research
    Labour --> Research
    Data --> Research
    Research --> Capability

    Capability --> Product
    Product --> UseValue
    Product --> Revenue
    Product --> Externality
    Product --> AdminGrowth

    Revenue --> LabPower
    UseValue -. only partly captured .-> Revenue
    Externality -. often excluded from .-> Revenue
    AdminGrowth --> Revenue

    PublicChoice --> Rules
    Externality --> PublicChoice
    UseValue --> PublicChoice
    Labour --> PublicChoice

    Rules --> Finance
    Rules --> LabPower
    Rules --> Data
    Rules --> Infrastructure
    LabPower --> Rules

The key correction

The paper’s route is approximately:

$$ C\rightarrow Y\rightarrow\text{investment}\rightarrow C. $$

The rewrite separates:

$$ \text{use value}, \quad \text{revenue}, \quad \text{financing}, \quad \text{organisational control}. $$

A system may generate substantial social usefulness without the AI firm capturing much revenue.

A firm may capture substantial revenue without generating equivalent social usefulness.

And a firm may attract enormous financing without current revenue because investors and states believe future capability will confer strategic power.

The latter is particularly important for frontier AI. The paper itself notes that lab expenditures are largely financed through outside investment and expectations rather than current revenue. [bol.com]

That means the operative loop may be:

$$ \text{capability demonstration} \rightarrow \text{belief in future dominance} \rightarrow \text{financing} \rightarrow \text{material expansion} \rightarrow \text{next capability demonstration}. $$

That is a sociotechnical expectations loop, not merely an economic-output loop.

And because financing gives organisations the ability to shape regulation, infrastructure, and standards, power feeds back into the rules governing future accumulation.

Diagram 5 — The whole cursed machine

This is the integrated version: technical recursion embedded inside institutional recursion.

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    subgraph Meaning["Meaning and legitimacy"]
        Metrics["Metrics and benchmarks"]
        Narrative["Narratives of progress<br/>necessity and inevitability"]
        Value["Contested human values<br/>what should count as improvement"]
    end

    subgraph Power["Power and institutions"]
        Owners["Owners, executives and states"]
        Rules["Law, governance and access"]
        Allocation["Resource and agenda allocation"]
        Workers["Researchers, workers and communities"]
        Resistance["Negotiation, refusal and contestation"]
    end

    subgraph Production["Technical production"]
        Inputs["Compute, energy, data and infrastructure"]
        Candidates["Candidate discoveries and outputs"]
        Verification["Verification, judgement and integration"]
        Systems["Deployed AI systems"]
    end

    subgraph Consequences["Consequences"]
        Useful["Useful knowledge and services"]
        Revenue["Revenue and financing"]
        Admin["Administrative output and control"]
        Costs["Externalised costs and displaced burdens"]
        Freedom["Human freedom and collective capacity"]
    end

    Systems --> Metrics
    Metrics --> Narrative
    Narrative --> Owners
    Value --> Allocation

    Owners --> Allocation
    Owners --> Rules
    Rules --> Allocation
    Workers --> Candidates
    Workers --> Verification
    Allocation --> Inputs
    Allocation --> Workers

    Inputs --> Candidates
    Systems --> Candidates
    Candidates --> Verification
    Verification --> Systems

    Systems --> Useful
    Systems --> Admin
    Systems --> Costs

    Useful --> Revenue
    Admin --> Revenue
    Revenue --> Owners

    Useful --> Freedom
    Admin -. may narrow .-> Freedom
    Costs -. may reduce .-> Freedom

    Costs --> Resistance
    Admin --> Resistance
    Workers --> Resistance
    Resistance --> Rules
    Resistance --> Value

    Narrative -. shapes .-> Metrics
    Metrics -. disciplines .-> Workers
    Candidates -. congestion .-> Verification

This diagram contains two recursive systems.

Technical recursion

$$ \text{systems} \rightarrow \text{candidate discoveries} \rightarrow \text{verification} \rightarrow \text{better systems}. $$

Institutional recursion

$$ \text{capability narrative} \rightarrow \text{authority and financing} \rightarrow \text{resource allocation} \rightarrow \text{capability demonstration} \rightarrow \text{stronger narrative}. $$

The paper concentrates on the first and adds parts of the second as an “economic feedback loop.” A Graeberian analysis would say the two cannot be cleanly separated.

The technical loop is powered, selected, measured, and constrained by the institutional loop. The institutional loop, in turn, gains legitimacy from the technical loop.

Why their boxes triggered your scepticism

I think your reaction came from four graphical tricks—not necessarily deliberate tricks, but tricks nonetheless.

1. Equal-looking boxes imply equal ontological solidity

“Compute” is physically measurable.

“Capability” is an index constructed from selected evaluations.

“Economic output” is an accounting convention.

“R&D effort” is a latent aggregate.

Yet all appear as equivalent rectangular objects.

2. Arrows conceal institutions

An arrow such as:

$$ C\rightarrow\dot A $$

sounds mechanical.

But the real path is closer to:

$$ \begin{aligned} \text{model output} &\rightarrow \text{human interpretation}\\ &\rightarrow \text{experiment selection}\\ &\rightarrow \text{resource approval}\\ &\rightarrow \text{experimental execution}\\ &\rightarrow \text{verification}\\ &\rightarrow \text{integration}\\ &\rightarrow \text{recognised improvement}. \end{aligned} $$

Every step contains contestable judgement and institutional power.

3. Negative feedback is relegated to prose

Positive feedback receives visible arrows.

Complexity, congestion, conflict, verification burdens, organisational dysfunction, resistance, and externalities mostly receive caveats.

What is drawn feels structural. What is written beneath it feels incidental.

4. The diagram has no subject

Who chooses?

Who benefits?

Who bears the cost?

Who is allowed to refuse?

Who defines improvement?

In the original diagrams, nobody does. Variables transform into other variables as if social arrangements were chemical reactions.

That is the deepest source of bullshit aroma.

Compressed Graeberian replacement

If I had to replace the paper’s entire graphical theory with one line, it would be:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    Imagination["Human imagination"] --> Institution["Institutions select and formalise"]
    Institution --> Resources["Power allocates resources"]
    Resources --> Technology["Technology expands selected capabilities"]
    Technology --> Administration["New possibilities for value and control"]
    Administration --> Institution
    Technology --> Consequences["Benefits, burdens and unintended consequences"]
    Consequences --> Contestation["People negotiate, resist and reorganise"]
    Contestation --> Imagination

The paper asks whether intelligence can recursively improve itself.

The Graeberian correction is:

There is no disembodied intelligence improving itself. There are people and institutions recursively reorganising the material world around particular formalised visions of intelligence—and being reorganised by the resulting machines.

That is why the original diagrams felt wrong. They were not necessarily making false statements.

They were making socially manufactured relationships look like natural machinery.

GPT-5.6:

🜁⊕ Addendum — faithful Mermaid reconstructions of the paper’s diagrams

Absolutely. Below are close Mermaid reproductions of Figures 1–6, preserving the authors’ own chosen boxes and causal arrows rather than applying the Graeberian corrections.

A technical note: the paper uses triangles to indicate accumulation of a flow into a stock—for example, algorithmic improvements $\dot A$ accumulating into algorithmic efficiency $A$. Mermaid does not have that exact stock-flow notation, so I represent accumulation using a dashed arrow labelled “accumulates into.” The causal/production-function arrows remain solid. [bol.com]

Figure 1 — Jones model of innovation

The original model says that R&D labour $L$ and the current stock of algorithmic efficiency $A$ affect the production of algorithmic improvements $\dot A$. Those improvements accumulate into $A$. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    L["R&D labour<br/>L"]
    A["Algorithmic efficiency<br/>A"]
    Adot["Algorithmic improvements<br/>Ȧ"]

    L --> Adot
    A --> Adot
    Adot -. "accumulates into" .-> A

The only feedback loop is:

$$ A\rightarrow\dot A\rightarrow A. $$

Figure 2 — Baseline model of RSI

The baseline RSI model adds AI capability $C$ and training compute $T$. Algorithmic efficiency and training compute produce capability; capability then contributes to further algorithmic improvement. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    L["R&D labour<br/>L"]
    A["Algorithmic efficiency<br/>A"]
    Adot["Algorithmic improvements<br/>Ȧ"]
    C["Capabilities<br/>C"]
    T["Training compute<br/>T"]

    L --> Adot
    A --> Adot
    C --> Adot

    A --> C
    T --> C

    Adot -. "accumulates into" .-> A

This contains two feedback paths:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    A["Algorithmic efficiency<br/>A"]
    C["Capabilities<br/>C"]
    Adot["Algorithmic improvements<br/>Ȧ"]

    A --> Adot
    A --> C
    C --> Adot
    Adot -. "accumulates into" .-> A

Direct self-feedback: $A\rightarrow\dot A$
Core RSI feedback: $A\rightarrow C\rightarrow\dot A$

Figure 3 — Model with human, compute, and data bottlenecks

This extends Figure 2 with:

experimental compute $E$;
inference compute $K$;
training compute $T$;
data $D$.

Human labour, capability, experimental compute, and inference compute contribute to algorithmic improvements. Algorithmic efficiency, training compute, and data contribute to capability. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    L["Human R&D labour<br/>L"]
    A["Algorithmic efficiency<br/>A"]
    Adot["Algorithmic improvements<br/>Ȧ"]
    C["Capabilities<br/>C"]

    E["Experimental compute<br/>E"]
    K["Inference compute<br/>K"]
    T["Training compute<br/>T"]
    D["Data<br/>D"]

    L --> Adot
    A --> Adot
    C --> Adot
    E --> Adot
    K --> Adot

    A --> C
    T --> C
    D --> C

    Adot -. "accumulates into" .-> A

A layout closer to the paper’s stacked presentation:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    L["Human R&D labour<br/>L"]

    subgraph Core[" "]
        direction LR
        A["Algorithmic efficiency<br/>A"]
        Adot["Algorithmic improvements<br/>Ȧ"]
        C["Capabilities<br/>C"]

        A --> Adot
        A --> C
        C --> Adot
        Adot -. "accumulates into" .-> A
    end

    subgraph Inputs["Complementary inputs"]
        direction LR
        E["Experimental compute<br/>E"]
        K["Inference compute<br/>K"]
        T["Training compute<br/>T"]
        D["Data<br/>D"]
    end

    L --> Adot
    E --> Adot
    K --> Adot
    T --> C
    D --> C

The extra boxes do not create new feedback loops because the model holds $E,K,T,D,$ and $L$ externally fixed. They instead affect how strong the existing arrows can remain.

Figure 4 — Narrow capabilities feed back; broad capabilities dangle

This model divides capability into:

$C_1$: narrow capability useful for AI R&D;
$C_2$: broad capability useful for producing economic output.

Algorithmic efficiency affects both kinds of capability, but only narrow capability feeds back into algorithmic improvement. Broad capability instead affects economic output $Y$. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    L["R&D labour<br/>L"]
    A["Algorithmic efficiency<br/>A"]
    Adot["Algorithmic improvements<br/>Ȧ"]

    C1["Narrow capabilities<br/>C₁"]
    C2["Broad capabilities<br/>C₂"]
    Y["Economic output<br/>Y"]

    L --> Adot
    A --> Adot
    C1 --> Adot

    A --> C1
    A --> C2
    C2 --> Y

    Adot -. "accumulates into" .-> A

The deliberately isolated paths are:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    A["Algorithmic efficiency<br/>A"]

    C1["Narrow capabilities<br/>C₁"]
    Adot["Algorithmic improvements<br/>Ȧ"]

    C2["Broad capabilities<br/>C₂"]
    Y["Economic output<br/>Y"]

    A --> C1 --> Adot
    Adot -. "accumulates into" .-> A

    A --> C2 --> Y

Thus:

the upper path closes the RSI loop;
the lower path “dangles” because $Y$ does not feed back into the technical system in this particular figure.

Figure 5 — AI helps improve one sub-algorithm

Here, overall algorithmic efficiency $A$ depends on two component efficiencies, $A_1$ and $A_2$.

Each component has:

corresponding human R&D labour;
a flow of improvements;
accumulation into its efficiency stock.

Capability $C$ directly helps improve only algorithm 2. Overall efficiency affects capability, closing a selective feedback loop through $A_2$. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    L1["Human R&D labour<br/>L₁"]
    A1["Algorithm 1 efficiency<br/>A₁"]
    A1dot["Improvements to algorithm 1<br/>Ȧ₁"]

    L2["Human R&D labour<br/>L₂"]
    A2["Algorithm 2 efficiency<br/>A₂"]
    A2dot["Improvements to algorithm 2<br/>Ȧ₂"]

    A["Overall efficiency<br/>A"]
    C["Capabilities<br/>C"]

    L1 --> A1dot
    A1 --> A1dot
    A1dot -. "accumulates into" .-> A1

    L2 --> A2dot
    A2 --> A2dot
    C --> A2dot
    A2dot -. "accumulates into" .-> A2

    A1 --> A
    A2 --> A

    A --> C

A paired layout closer to the figure:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    subgraph Algorithm1["Algorithm 1"]
        direction TB
        L1["Human R&D labour<br/>L₁"]
        A1dot["Improvements to algorithm 1<br/>Ȧ₁"]
        A1["Algorithm 1 efficiency<br/>A₁"]

        L1 --> A1dot
        A1 --> A1dot
        A1dot -. "accumulates into" .-> A1
    end

    subgraph Algorithm2["Algorithm 2"]
        direction TB
        L2["Human R&D labour<br/>L₂"]
        A2dot["Improvements to algorithm 2<br/>Ȧ₂"]
        A2["Algorithm 2 efficiency<br/>A₂"]

        L2 --> A2dot
        A2 --> A2dot
        A2dot -. "accumulates into" .-> A2
    end

    A["Overall efficiency<br/>A"]
    C["Capabilities<br/>C"]

    A1 --> A
    A2 --> A
    A --> C
    C --> A2dot

The AI-assisted loop is:

$$ A_2\rightarrow A\rightarrow C\rightarrow\dot A_2\rightarrow A_2. $$

Algorithm 1 can eventually become the bottleneck because capability does not directly accelerate $\dot A_1$.

Figure 6 — Model with economic feedback loops

This is the largest diagram. It adds economic output $Y$, generated partly by AI capabilities. Output drives investment or accumulation flows for:

experimental compute;
inference compute;
training compute;
data.

Those flows accumulate into stocks $E,K,T,$ and $D$, which feed back into algorithmic improvement or capability production. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    L["Human R&D labour<br/>L"]
    A["Algorithmic efficiency<br/>A"]
    Adot["Algorithmic improvements<br/>Ȧ"]
    C["Capabilities<br/>C"]
    Y["Economic output<br/>Y"]

    E["Experimental compute<br/>E"]
    K["Inference compute<br/>K"]
    T["Training compute<br/>T"]
    D["Data<br/>D"]

    Edot["Investment in experimental compute<br/>Ė"]
    Kdot["Investment in inference compute<br/>K̇"]
    Tdot["Investment in training compute<br/>Ṫ"]
    Ddot["Production or collection of data<br/>Ḋ"]

    L --> Adot
    A --> Adot
    C --> Adot
    E --> Adot
    K --> Adot

    A --> C
    T --> C
    D --> C

    C --> Y

    Y --> Edot
    Y --> Kdot
    Y --> Tdot
    Y --> Ddot

    Edot -. "accumulates into" .-> E
    Kdot -. "accumulates into" .-> K
    Tdot -. "accumulates into" .-> T
    Ddot -. "accumulates into" .-> D

    Adot -. "accumulates into" .-> A

A more compact reproduction:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    L["Human R&D labour<br/>L"]

    A["Algorithmic efficiency<br/>A"]
    C["Capabilities<br/>C"]
    Y["Economic output<br/>Y"]
    Adot["Algorithmic improvements<br/>Ȧ"]

    E["Experimental compute<br/>E"]
    K["Inference compute<br/>K"]
    T["Training compute<br/>T"]
    D["Data<br/>D"]

    Edot["Ė"]
    Kdot["K̇"]
    Tdot["Ṫ"]
    Ddot["Ḋ"]

    L --> Adot
    A --> Adot
    C --> Adot
    E --> Adot
    K --> Adot
    Adot -.-> A

    A --> C
    T --> C
    D --> C

    C --> Y

    Y --> Edot
    Y --> Kdot
    Y --> Tdot
    Y --> Ddot

    Edot -.-> E
    Kdot -.-> K
    Tdot -.-> T
    Ddot -.-> D

The main feedback families are easier to see when separated.

Core technical feedback

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    A["Algorithmic efficiency<br/>A"]
    C["Capabilities<br/>C"]
    Adot["Algorithmic improvements<br/>Ȧ"]

    A --> C --> Adot
    Adot -. "accumulates into" .-> A

Training-compute economic feedback

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    C["Capabilities<br/>C"]
    Y["Economic output<br/>Y"]
    Tdot["Training-compute investment<br/>Ṫ"]
    T["Training compute<br/>T"]

    C --> Y --> Tdot
    Tdot -. "accumulates into" .-> T
    T --> C

Data economic feedback

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    C["Capabilities<br/>C"]
    Y["Economic output<br/>Y"]
    Ddot["Data production or collection<br/>Ḋ"]
    D["Data<br/>D"]

    C --> Y --> Ddot
    Ddot -. "accumulates into" .-> D
    D --> C

Experimental-compute feedback

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    C["Capabilities<br/>C"]
    Y["Economic output<br/>Y"]
    Edot["Experimental-compute investment<br/>Ė"]
    E["Experimental compute<br/>E"]
    Adot["Algorithmic improvements<br/>Ȧ"]
    A["Algorithmic efficiency<br/>A"]

    C --> Y --> Edot
    Edot -. "accumulates into" .-> E
    E --> Adot
    Adot -. "accumulates into" .-> A
    A --> C

Inference-compute feedback

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart LR
    C["Capabilities<br/>C"]
    Y["Economic output<br/>Y"]
    Kdot["Inference-compute investment<br/>K̇"]
    K["Inference compute<br/>K"]
    Adot["Algorithmic improvements<br/>Ȧ"]
    A["Algorithmic efficiency<br/>A"]

    C --> Y --> Kdot
    Kdot -. "accumulates into" .-> K
    K --> Adot
    Adot -. "accumulates into" .-> A
    A --> C

One-page combined reproduction

This version combines the paper’s conceptual progression into a single Mermaid diagram. It is not one of the original figures, but it contains the principal nodes and relationships introduced across Figures 1–6. [bol.com]

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ccc"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    subgraph Research["AI research production"]
        L["Human R&D labour<br/>L"]
        E["Experimental compute<br/>E"]
        K["Inference compute<br/>K"]
        C1["Narrow AI-R&D capability<br/>C₁"]
        Adot["Algorithmic improvements<br/>Ȧ"]

        L --> Adot
        E --> Adot
        K --> Adot
        C1 --> Adot
    end

    subgraph Capability["Capability production"]
        A["Algorithmic efficiency<br/>A"]
        T["Training compute<br/>T"]
        D["Data<br/>D"]
        C2["Broad capability<br/>C₂"]

        A --> C1
        A --> C2
        T --> C1
        T --> C2
        D --> C1
        D --> C2
    end

    Adot -. "accumulates into" .-> A
    A --> Adot

    subgraph Economy["Economic feedback"]
        Y["Economic output<br/>Y"]
        Edot["Ė"]
        Kdot["K̇"]
        Tdot["Ṫ"]
        Ddot["Ḋ"]

        C2 --> Y
        Y --> Edot
        Y --> Kdot
        Y --> Tdot
        Y --> Ddot
    end

    Edot -. "accumulates into" .-> E
    Kdot -. "accumulates into" .-> K
    Tdot -. "accumulates into" .-> T
    Ddot -. "accumulates into" .-> D

The useful contrast

We now have two diagram sets:

The paper’s diagrams: a technical-economic production system made from scalar inputs, outputs, and elasticities.
The Graeberian diagrams: a sociotechnical system made from institutions, narratives, judgement, power, material resources, verification, contestation, and competing concepts of value.

Placing them side by side makes the critique spectacularly legible:

The original arrows are not necessarily false. They are compressed tunnels through which an enormous amount of social machinery has been made to disappear.

Paper-to-Graeber diagram crosswalk

Paper reference	Paper diagram	Closest Graeberian equivalent developed here
Figure 1	Jones Model of Innovation	Diagram 1 — “AI improves AI,” except society is doing nearly all the verbs
Figure 2	Baseline Model of RSI	Diagram 1 — “AI improves AI,” except society is doing nearly all the verbs
Figure 3	Model with Bottlenecks	Diagram 2 — The Missing Verification and Bullshit-Production Loop
Figure 4	Narrow Capabilities Feed Back; Broad Capabilities Dangle	Diagram 3 — Narrow Capability Does Not “Dangle”; Institutions Decide Where It Goes
Figure 5	AI Helps One Subalgorithm	Diagram 2 — The Missing Verification and Bullshit-Production Loop
Figure 6	Model with Economic Feedback Loops	Diagram 4 — The Economic Feedback Loop, with Appropriation and Belief Restored
Figures 1–6 combined	The paper’s complete recursive-improvement framework	Diagram 5 — The Whole Cursed Machine
Conceptual summary	No direct paper equivalent	Compressed Graeberian Replacement — Human imagination → institutions → resources → technology → consequences → contestation

Important qualification

The relationship is not perfectly one-to-one:

Graeber Diagram 1 expands the paper’s basic technical loop in Figures 1–2.
Graeber Diagram 2 cuts across Figures 3 and 5 by exposing the verification and congestion processes hidden inside their arrows.
Graeber Diagram 3 is a direct critique and expansion of Figure 4.
Graeber Diagram 4 is a direct critique and expansion of Figure 6.
Graeber Diagram 5 embeds the paper’s entire technical system inside institutional and political recursion.

The cleanest condensed mapping is therefore:

$$ \boxed{ \text{Paper Figures 1–2} \rightarrow \text{Graeber Diagram 1} } $$

$$ \boxed{ \text{Paper Figures 3 and 5} \rightarrow \text{Graeber Diagram 2} } $$

$$ \boxed{ \text{Paper Figure 4} \rightarrow \text{Graeber Diagram 3} } $$

$$ \boxed{ \text{Paper Figure 6} \rightarrow \text{Graeber Diagram 4} } $$

$$ \boxed{ \text{Paper’s entire framework} \rightarrow \text{Graeber Diagram 5} } $$

Some mermaid diagrams generated from this...

the_economics_of_recursive_self_improvement.mmd:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ddd"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    subgraph Material
        Experiments
        Data
        Compute
        Verification
        Systems
    end

    subgraph Narrative
        Metric
        Progress
        Story
    end

    subgraph Institution
        Authority
        Labour
        Management
        Ownership
    end

    Systems --> Metric
    Metric --> Progress --> Story
    Story --> Authority

    Authority --> Management & Labour
    Management --> Compute & Data
    Ownership --> Data & Labour

    Labour & Systems & Compute & Data --> Experiments

    Experiments -- increases burden --> Verification --> Systems & Progress
    Labour & Compute --> Verification

    Management -. may constrain .-> Labour
    Ownership -. limits diffusion .-> Progress

the_economics_of_rsi_graeber.mmd:

---
config:
    look: "neo"
    theme: "base"
    fontFamily: Avenir Next, Avenir, system-ui, Helvetica Neue, sans-serif
    themeVariables:
        clusterBkg: "#ffffff"
        clusterBorder: "#d9e8f5"
        lineColor: "#ddd"
        primaryBorderColor: "#c5d9ed"
        primaryColor: "#f0f7ff"
        secondaryColor: "#fff"
        primaryTextColor: "#1a2332"
        titleColor: "#1B4965"
---
flowchart TB
    Imagination["Human imagination"] --> Institution["Institutions select and formalise"]
    Institution --> Resources["Power allocates resources"]
    Resources --> Technology["Technology expands selected capabilities"]
    Technology --> Administration["New possibilities for value and control"]
    Administration --> Institution
    Technology --> Consequences["Benefits, burdens and unintended consequences"]
    Consequences --> Contestation["People negotiate, resist and reorganise"]
    Contestation --> Imagination

ianchanning/the_economics_of_recursive_self_improvement.md

🜁⊕ Acorn received: legibility before belief

Cold-silicon verdict

1. What the diagrams actually are

$$ \varepsilon_{C,A}

2. What they get genuinely right

A. They separate automation from acceleration

B. They expose bottlenecks rather than assuming frictionless software recursion

C. Narrow versus broad capability is an excellent distinction

D. They admit the empirical holes

3. Where the diagrams smuggle assumptions into the room

A. The graph topology is treated as organisation, but it is actually the central hypothesis

B. Several nodes are implausibly scalar

C. Their arrows are local derivatives, not durable mechanisms

4. The dodgiest economic assumption: cost shares equal elasticities

5. The balanced-growth manoeuvre is mathematically neat and empirically fragile

$$ \frac{\varepsilon_{\dot A,R}} {1-\varepsilon_{\dot A,A}}

6. “Algorithmic efficiency” is partly a residual, and residuals are hungry ghosts

7. The 15% threshold: algebraically clean, epistemically feral

The first factor

The third factor

The missing factor

8. The claimed 9% estimate is the paper’s weakest passage

9. Figure-by-figure bullshit detector

Figure 1: Jones innovation model

Figure 2: baseline RSI model

Figure 3: bottlenecks

Figure 4: narrow capabilities feed back; broad capabilities dangle

Figure 5: AI helps one subalgorithm

Figure 6: economic feedback loops

10. What a stronger diagram would need

1. Discovery

2. Verification

3. Integration

4. Material capacity

5. Governance and coordination

11. The right interpretation of the paper

$$ \text{RSI loop gain}

Final judgement

User:

GPT-5.6:

User:

GPT-5.6:

🜁⊕ The Graeber lens

1. He would attack the paper’s ontology before its arithmetic

2. The paper mistakes an institution for a machine

3. “Economic output” would set off every alarm bell

4. The hidden subject of the paper is not AI—it is capital allocation

5. He would question their concept of “intelligence”

6. The narrow/broad distinction would interest him—but not satisfy him

7. He would see the data requests as a bid for bureaucratic legibility

8. The “political and social barriers” section would infuriate him

9. What he might nevertheless admire

The authors expose their assumptions

They reject technological inevitability

They identify empirical ignorance

The Graeberian rewrite of the central question

My compressed verdict

User:

GPT-5.6:

🜁⊕ Yes. Let’s put the missing society back into the machine.

Diagram 1 — “AI improves AI,” except society is doing nearly all the verbs

What changed?

Diagram 2 — The missing verification and bullshit-production loop

The heresy contained here

Productive loop

Bureaucratic amplification loop

Diagram 3 — Narrow capability does not “dangle”; institutions decide where it goes

Why this is better than “narrow versus broad”

Diagram 4 — The economic feedback loop, with appropriation and belief restored

The key correction

Diagram 5 — The whole cursed machine

Technical recursion

Institutional recursion

Why their boxes triggered your scepticism

1. Equal-looking boxes imply equal ontological solidity

2. Arrows conceal institutions

3. Negative feedback is relegated to prose

4. The diagram has no subject

Compressed Graeberian replacement