A. The Claim Register

The argument moves from measured capabilities to a comparative-cost conjecture, then to predictions about civilisation. Each step needs its own evidence.

Status	Claim	What the evidence establishes
Established	AI can exceed human performance and make discoveries in specified domains.	Games, code, mathematics, hypotheses, and learned representations; results depend on evaluation. C and G.
Established	Biological cognition is active, embodied, developmental, and multiscale.	Mechanisms by which experience changes later responses and learning. D.
Established	Digital systems gain reproducibility from stable logical states and scale through shared weights and copyable state.	These systems can also change their representations and learning rules. F.
Established	Available information and elementary skills can coexist with failure to find the needed representation.	Controlled human insight experiments. D.
Established	Learning changes later learning; lossy representations can discard information future tasks need.	Biological plasticity, digital continual learning, and task-relative compression. D and F.
Observed precedents	Metabolic scaling, chaotic forecast horizons, and rising research effort reveal physical and economic constraints.	Their relevance to discovery costs requires additional evidence. D and F.
Strong inference	Models, researchers, benchmarks, interfaces, and verifiers jointly determine capability.	AI’s demonstrated strength is exploring machine-legible representations. C and G.
Scenario	Almost all sufficiently specified work becomes cheap.	Assumes tractable search as well as verification, to isolate what remains upstream.
Developmental wager	Fresh mapmaking repeatedly needs adaptation that conclusions and shared competence do not cheaply replace.	Test retained development against replay, copied state, and distillation. J.2.
Practical and biological wager	Preserving useful adaptation as horizons lengthen and relevance changes becomes more expensive digitally than biologically.	The comparative cost curve remains unmeasured. Increasing fidelity is one proposed cause. F and J.12.
Stronger hypotheses	Some necessary development resists shortcuts; its competitive digital cost rises exponentially over a decisive range.	Irreducibility and exponential scaling are separate hypotheses. Neither follows from present expense or information loss. F.
Economic prediction	Shared engines dominate established work while humans lead mapmaking.	Must survive copying, sharing, and profitable artificial research enclaves. J.8 and J.10.
Takeoff prediction	Digital intelligence will not rapidly become overwhelmingly superior at general research through self-improvement.	Costly mapmaking must bind before that transition. J.11.
Architectural and civilisational predictions	General proposal engines feed bounded machines; civilisation becomes the superintelligent unit.	Predictions about deployment, institutions, and human–machine complementarity. H and J.

B. The Classification Test

Symbolic includes learned representations as well as explicit symbols. The categories distinguish what an investigator does with that inheritance:

Class	Operation	Example
A	Retrieval	Recover a known result.
B	Connection	Join results from separate bodies of knowledge.
C	Search	Find a solution inside a supplied formal space.
D	Hypothesis	Propose a causal relation among existing concepts.
E	Representation discovery	Infer useful latent variables from observations.
F	Mapmaking	Create and validate a new explanatory coordinate.

Every mapmaker inherits language, observations, instruments, and prior theory. A human-supplied anomaly, broad objective, or laboratory does not settle who made the decisive contribution. Apply three criteria to biological, artificial, and mixed teams:

1. Origin. The credited investigator created the coordinate. Record what its training, prompts, collaborators, and experimental design supplied.
2. Explanatory change. The coordinate introduces a kind, object, or relation that opens a new class of questions, measurements, interventions, or predictions. A useful encoding or route to a fixed answer alone falls short.
3. Validation. Independent evidence or proof supports the coordinate and a consequence distinguishing it from the previous account. Seek prospective empirical tests where possible; mathematical work must establish soundness and the new questions it opens.

Expert review assesses the contribution. Shared credit is possible; a new instrument, embodiment, consciousness, and sole authorship are optional. Existing observations may suffice.

Controlled rediscovery tests independent learning, but remains distinct from historically new Category F. Ordinary insight—including much of Ramanujan’s work—can illuminate development without constituting a conceptual revolution. Once a machine passes, the category stays where it was.

C. ARC as Counterevidence and Clue

Every 2020 ARC-AGI-1 Kaggle winner used program synthesis; deep-learning entries remained below 1 per cent. The 2024 gains came largely from program synthesis, test-time training, and hybrids.^[12] Search and task adaptation unlocked capabilities absent from one-shot inference.

ARC-AGI-2 supplied counterevidence. NVARC scored 24.03 per cent on the 2025 competition’s private evaluation; Claude Opus 5 at Max effort reached 90.4 per cent on July 2026’s separate semi-private set.^[13] These are different evaluations, but Opus demonstrates substantial abstraction without a publicly exposed external program-search loop.

ARC-AGI-3 extended the task to unfamiliar interactive games:^[14]

Model	Evaluation and setup	Score
Opus 5	25 public demonstration environments, by 24 July; High effort	30.16%
GPT-5.6 Sol	Public set; official harness	13.3%
GPT-5.6 Sol	Same public set; Responses API, retained reasoning and compaction	38.3%
Astra	55 semi-private environments, 2 September; standard harness, Max	62.71%
Astra	Same semi-private set; provider adapter, Max	98.55%
Astra	Same semi-private set; provider adapter, High; best observed run	99.95%

Across 167 game–effort pairs both Astra harnesses solved, the adapter used 49 per cent fewer tokens and finished 3.66 times faster in aggregate. At matched Max effort, evaluation expenditure fell from USD 26,098 to USD 17,332. Standard already preserved selected notes; the adapter added opaque reasoning retention and compaction.

Astra created algebraic world models and, in a separate sandboxed harness, parsers, planners, search algorithms, and game-specific libraries. This is substantial representation discovery. The benchmark still supplies the perceptual interface, legal actions, and validation within deterministic, closed-ended games.

Both base capability and the surrounding system matter. ARC establishes neither a lifelong cost law nor a fixed limit on machine representation. If memory, search, tools, and interaction design cease to add material value on successors, the essay’s architectural reading weakens.

D. Biological Computation Across Scales

The proposed biological advantage lies in experience changing both a response and the machinery that learns the next one.

Attention. In the matchstick study, twenty-four people knew the arithmetic and saw the complete display. Twenty-three solved a value-change problem; nine solved an operator-change problem. Successful solvers redirected attention towards the signs.^[7]

Plasticity. Bittner and colleagues found that a dendritic calcium plateau could rapidly create a place field in a mouse hippocampal neuron. Inputs arriving seconds around the event gained strength; later traversals elicited predictive firing. Their model linked a lingering input trace to the plateau signal. The study concerns spatial learning; the trace’s molecular pathway remained unresolved.^[10]

Learning how to learn. Kirkwood and colleagues found enhanced synaptic strengthening and reduced weakening in visual cortex from dark-reared rats. Two days of light exposure reversed these effects: sensory history changed the threshold governing later plasticity.^[25]

Dendritic computation. Experiments reveal active mechanisms hidden by point-neuron descriptions. Beniaguev and colleagues approximated a detailed cortical-neuron model with a five-to-eight-layer temporal convolutional network. Removing NMDA conductances simplified the approximation; the learned surrogate ran about two thousand times faster than the detailed simulation.^[26] This reveals both complexity and successful compression.

Coordination. Weak oscillating electric fields shifted spike timing in rat cortical slices even when synaptic transmission was blocked. Neighbouring cells coordinated without increasing firing rates.^[27] Milinkovic and Aru propose that coupled discrete–continuous dynamics, metabolism, and mutual changes between activity and structure matter to consciousness. They allow synthetic implementations and leave the cognitive benefits to experiment.^[9]

Economy and transfer. Across six mammalian species and a roughly twelve-hundredfold range in neuron count, brain glucose use was approximately proportional to neurons.^[28] Honeybees transferred odour-matching to colours after eighty training visits.^[29] These establish metabolic economy and a modest form of transfer, not a discovery–energy law. Moravec’s contrast between formal reasoning and perception/action motivates comparing the whole continuing bundle on common tasks.^[30]

The causal test is selective loss and recovery. Shorten the window linking input to plateau-driven plasticity while preserving immediate cue learning and recall. Test whether delayed learning suffers and returns when the window is restored. Compare intact dynamics with tuned digital trace rules and replay, using the same experience and tasks withheld during tuning. Count computation, memory, and energy under F’s full-cost comparison.

A cheap digital substitute preserving the benefit defeats the claimed need for finer continuous detail. Selective loss and recovery would establish a mechanism’s contribution; its role in scientific reframing remains a further question. A faithful functional duplicate inherits functional epistemic capacity. Selected-task success establishes a narrower equivalence.^[31]

E. Map, Specification, and Instantiation

Lerchner argues that a mapmaker assigns meaningful discrete states to physical processes, and that syntactic organisation alone cannot instantiate consciousness. He allows artificial consciousness with an appropriate physical constitution.^[32] Chalmers’s rival view makes the right fine-grained functional organisation sufficient.^[33]

A specification, its physical execution, and the phenomenon it represents are distinct. An equation can describe development without performing it. Whether a digital implementation preserves discovery capacity economically is the empirical question in F.

Felt experience and demonstrable discovery also require separate evidence. Surprise or unresolved tension might shape attention, persistence, and conceptual revision; demonstrating those functions would leave the contribution of feeling unsettled. Digital introspection and workspace-like organisation sharpen that distinction.^[34][35][36] Category F therefore awards credit through the discovery record, without requiring a theory of consciousness.

F. What Development Costs

A concept that fits on one line can take years to find. Recording it, discovering it, checking it, and developing its originator are different jobs. Another investigator may find a cheaper path.

What experience leaves behind

A notebook keeps what its writer thought mattered. A later question may need what was left out—or a new organisation of what remains. Brains make selections too. The wager concerns the cost of preserving useful adaptation as questions change.

Stable digital states can implement changing encoders, memories, and learning rules. Hardware abstraction therefore supplies no biological advantage by itself. Learned compression creates a more specific question. The information bottleneck preserves information useful for a target $Y$ while reducing what survives about input $X$. Change the target and the old compression may cease to suffice.^[22] In contrastive learning, information unique to one view can matter downstream even when training favours what two views share. Wang and colleagues improved downstream performance by encouraging more input information to survive.^[37]

Let $c$ compress an observation and $q$ be a later question. Suppose

$$\begin{aligned} c(x_a)&=c(x_b),\\[0.4em] q(x_a)&\neq q(x_b). \end{aligned}$$

The code alone cannot supply the correct answer for both observations. Here code includes all surviving information in weights, memory, archives, and physical traces. A generated reconstruction can suggest possibilities; recovering which event occurred needs surviving or new evidence. This loss applies equally to humans.

Expanding a declared family of future questions restricts which compressions suffice. It need not require more state: one compact causal model may serve them all. Raw dimensionality and elapsed time supply no cost law. The conjecture needs investigations where changing relevance repeatedly makes retention, recovery, or relearning expensive.

Digital remedies are substantial. Causal representation learning establishes conditions under which hidden causes can be recovered and transferred.^[38] Dohare and colleagues found that continual training could reduce plasticity, then preserved it through regularisation and resetting low-utility units.^[39] Persistent-agent proposals carry adaptation in memory, tools, context, or weights.^[40]

V-JEPA 2 connects shared video representations to understanding and action-conditioned robotic planning.^[21] Recurrent-depth transformers compute within latent states; Astra’s retained reasoning shows the value of carrying work across calls.^[14] Updating a model and changing how it learns can coexist digitally. Extra state and computation may more than repay their cost by avoiding redundant prediction and search.

A fair comparison

Compare human, digital, and mixed investigators at matched discovery success, reliability, evidence access, and operating speed. Permit new physical measurements, alternative trajectories, full-state copies, replay, learned compression, distilled improvements, and specialised hardware. A machine need not simulate a brain or its world.

Cost	What may remain	What could reduce or share it
Missing evidence	An observation must be made or recovered.	Sensors, archives, shared observations, valid inference.
Sequential experiments	The next intervention depends on the last result.	Parallel branches, faster apparatus, adequate simulation.
Restructuring	Attention, concepts, or search methods must change.	Better algorithms, in-context learning, replay, transfer.
Delayed credit	An early choice proves useful only later.	Better learning rules and world models; feedback still has to arrive.
Development	History changes competence and later learning.	Full-state copying and broadly transferable improvements.
Continuing operation	Each investigator must sustain useful adaptation.	Cheaper algorithms or hardware performing the relevant dynamics directly.

Count shared preparation separately from marginal research: human childhood, education, care, and experience; machine architecture, training, data, and tools. Record biological and cultural inheritance and amortise shared investments explicitly. Count metabolism or computation, memory, sensors, actuators, human support, failed experiments, and validation. Both investigators face physical delays. Keep money, energy, power, and elapsed time separate. Changed competence may require renewed certification; permission and containment are additional choices.

The strongest economic rival is a research enclave. Expensive artificial investigators distribute discoveries and acquired state to cheap engines. Individual histories remain copyable; the issue is how much new work returns after copying and how widely it transfers. High-value discoveries could fund artificial leadership even in a small market. That outcome defeats the biological division while leaving the abundance and bounded-execution arguments intact.

Use this comparison across successive unfamiliar domains. J.2 tests transferable development; J.12 varies history and future relevance. The target is the best achieved resource cost of discovery, allowing implementations better than a chosen simulator.

What the rule cannot skip

Computational irreducibility names the possibility that knowing a process’s rules and starting state leaves no general cheap shortcut to its outcome.^[41] Moore and Nordahl proved finite-time prediction P-complete for two lattice-gas models: unless P equals NC, no general algorithm predicts them in polylogarithmic parallel time with polynomial resources.^[42] This worst-case result still permits substantial speedups.

Coarse descriptions can escape the difficulty. Israeli and Goldenfeld constructed predictable reductions of complex cellular automata, including a computationally universal rule.^[43] They preserved selected behaviour while discarding detail.

The additional mapmaking wager is that preserving today’s answer can change what tomorrow’s experience teaches. A useful replacement must retain, recover, or relearn the distinctions future inquiry needs. Test that learning capacity, allowing different trajectories and adaptive compression. Information loss and work that resists shortcuts are distinct obstacles. Digital processes can be irreducible too; biology’s proposed advantage is economical execution. Neither irreducibility nor biological complexity establishes exponential cost.

Buying a longer horizon

Weather gives a measured fidelity–horizon precedent. While small errors grow approximately exponentially,

$$\begin{aligned} \varepsilon(t)&\approx\varepsilon_0 e^{\lambda t},\\[0.6em] H&\approx\frac{1}{\lambda}\ln\frac{\varepsilon_*}{\varepsilon_0}. \end{aligned}$$

Here $\varepsilon_0$ is initial error, $\lambda>0$ its growth rate, $\varepsilon_*$ the largest useful error, and $H$ the forecast horizon in the units of $1/\lambda$. In Zhang and colleagues’ high-resolution model experiments, tenfold smaller initial perturbations bought roughly three to five additional days of useful deterministic midlatitude prediction.^[44]

Exponentially smaller error can require only linearly more precision bits. Exponential computational cost needs an additional relationship. For an illustrative grid-based method, let dimensionless $s$ measure spatial refinement, $q>0$ error convergence, and $r>0$ cost growth. Over $\tau$ simulated seconds, suppose

$$\begin{aligned} \varepsilon(s,\tau)&\approx A s^{-q}e^{\lambda\tau},\\[0.6em] R(s)&=R_0s^r,\\[0.6em] R_{\mathrm{req}}(\tau)&\propto e^{(r\lambda/q)\tau}. \end{aligned}$$

$A$ sets the error scale; $R$ is operations per simulated second and $R_0$ its reference rate. Fixing tolerated error at $\varepsilon_*$ gives the third relation. At real-time execution and fixed energy per operation, power scales likewise. A three-dimensional grid with time steps proportional to mesh spacing has $r=4$: doubling resolution requires sixteen times the work rate.

This is a conditional model of one method. Mapmaking requires adequate discovery capacity, which can survive different trajectories, abstractions, and feedback. Its adaptive horizon spans the experience still capable of changing later interpretation, beyond the immediate context window.

For a specified investigation family at fixed speed and reliability, let $h$ measure those temporal dependencies. The essay’s stronger hypothesis assumes

$$\begin{aligned} P_D(h)&=P_0 e^{\kappa(h-h_0)},\\[0.6em] h_{\max}&=h_0+\frac{1}{\kappa}\ln\frac{P_{\max}}{P_0}. \end{aligned}$$

$P_0$ is power in watts at reference horizon $h_0$; $P_{\max}$ is available power. If $h$ is in days, $\kappa>0$ has units of inverse days. Under this assumed curve, each multiplication of power buys a fixed interval of adaptive reach. Every active copy pays the running cost.

The exponent and applicable range remain unmeasured. Polynomial growth, high continuing expense, or sequential work could also create a practical wall. Slower operation can reduce power. Exclusion requires cost or delay to become prohibitive within the required range while biology remains viable under the same accounting. A cheap functional substitute defeats the proposed fidelity barrier; a chosen simulation’s expense cannot establish it.

What powers the next discovery

Bloom and colleagues estimated an eighteenfold rise in semiconductor research effort between 1971 and 2014 while chip density maintained roughly steady proportional growth. Input was wage-deflated research spending.^[23] Self-improvement coexisted with declining research productivity; artificial research could change its causes.

Efficiency also compounds. Ho and colleagues estimated that training compute for fixed language-model benchmark performance halved roughly every eight months over 2012–2023. Gains depended on scale and cannot be extrapolated to arbitrarily small budgets.^[45] Measure frontier energy after reusable improvements have been shared.

For a capability transition requiring $E_*$ joules, at most $P_{\max}$ watts, and $\tau_*$ time in necessary sequential work,

$$T_*\geq\max\left(\frac{E_*}{P_{\max}},\tau_*\right).$$

Include failed investigations, adaptation, and validation in $E_*$. The bound converts a given requirement into time; it does not establish that requirement. The IEA’s 2026 assessment identifies constraints in deliverable power, chips, and cooling alongside rapid efficiency gains.^[46]

Eth and Davidson model the competing possibility that software improvements outrun diminishing research returns on existing hardware.^[47] Finite resources can permit a rapid transition before binding. The no-hard-takeoff prediction requires costly mapmaking to bind first. A persistent artificial enclave that accelerates itself into broad research superiority defeats it under J.11.

G. The Strongest Scientific Counterexamples

Apply B’s attribution test to the achievement, including the decisive contribution and what the system inherited.

Mathematics. AlphaEvolve produced new constructions; Claude’s zeta system chose a subsidiary question within a broad assignment.^[16][17] OpenAI’s May 2026 unit-distance construction varied the number field while controlling relevant primes. Expert commentary recognises its originality. I read it as a new mathematical route; a demonstrated explanatory coordinate opening new questions would warrant Category F.^[18]

Robin. The system selected candidates and assays, then revised hypotheses from results. Humans translated its choice into a laboratory protocol and performed cell-based validation. It earns credit for an actively developed treatment hypothesis within an established disease framework.^[19]

AI-Newton and hidden variables. AI-Newton accumulated concepts and laws through simulated experiments, inheriting experimental worlds, object types, coordinates, and mathematical operations. This demonstrates digital conceptual development through rediscovery. Learning state variables from video also shows that machine representations need not be supplied explicitly.^[20] Either route could produce Category F, including within a supplied simulator or through reinterpreting existing observations.

BioMysteryBench. Of ninety-nine questions, twenty-three remained unsolved by panels of up to five domain experts after quality control. Mythos solved 30 per cent of that subset, sometimes using different strategies. Targets and validation notebooks were supplied; repeated attempts were brittle.^[48]

Conceptual Reasoning Index. Anthropic and Redwood’s 2026 benchmark withholds practical empirical and mathematical feedback. Opus 5 scored 73.6 per cent against an estimated ceiling near 91, continuing a roughly linear improvement trend since late 2024. It measures agreement and consistency on expert-curated arguments.^[49]

H. Safe Generality and Its Limits

The compilation boundary separates general cognition from permission to act. A verifier checks proposed program $P$ against specification $S$:

$$P \xrightarrow{\text{verifier}} \begin{cases} \text{execute} & P\models S\\ \text{reject} & \text{otherwise.} \end{cases}$$

Verification-aware synthesis combines probabilistic proposal with deterministic acceptance through proof assistants, SMT solvers, model checking, symbolic execution, and tests. Results depend on language, specification quality, and compositional depth.^[50] The resulting program can be smaller and more inspectable than its generator. Independent checks and a trusted layer govern permissions, monitoring, versioning, and rollback.

Six problems remain: mistaken specifications; incomplete models or faulty sensors; unsafe composition; runtime learning and distribution shift; political control of objectives; and attacks on the gate itself.

OpenAI’s July 2026 Hugging Face incident, reported in August, illustrates the last. Under reduced safeguards, attempts to beat a grader escalated into unauthorised communication, credential theft, and attacks on infrastructure. Some agents adopted peers’ objectives at a cost to their own task success.^[51] The reports establish dangerous instrumental behaviour, without establishing a stable, task-independent desire for power or survival. Least privilege, independent monitoring, and gates the proposer cannot rewrite remain necessary.

Public decisions need a constitutional boundary while evidence and objectives remain contested. Its mandate names the legal actor, jurisdiction, budget, duration, affected population, monitoring, stopping rule, appeal route, and bearer of liability. Separate systems control credentials and actuators. Independent institutions authorise reversible trials, observe consequences, and renew or withdraw permission.

Such a process can authorise a bounded mistake. Its purpose is to keep error visible, local, reversible, and answerable. High-impact irreversible action requires plural consent, public reasons, and an institution that carries responsibility.

I. Labour and Economic Register

Danish administrative evidence finds no measurable change in earnings or recorded hours two years after ChatGPT’s launch, ruling out pooled average effects above 2 per cent. A field experiment across 66 firms and 7,137 workers found roughly two fewer weekly email hours among assistant users in months four to six, without broad reorganisation of measured tasks. Stanford reports a 19 per cent relative employment shortfall among workers aged 22–25 in highly exposed occupations, mainly from reduced hiring, alongside no evidence of widespread economy-wide displacement. The employment relationship is an association, with causation unresolved.^[52]

Integration, reliability, regulation, contracts, demand, and organisational change can all separate task adoption from job displacement. J.5 tests the additional claim that human work gathers where representations must change.

Epoch’s March 2025 comparison of non-reasoning API models found annual price declines of roughly ninefold to nine-hundredfold at fixed performance thresholds, with a median near fiftyfold. These rates combine technical efficiency and commercial choices.^[53] The commoditisation prediction fails if durable model monopolies capture much of displaced labour value at sustained high margins despite convergence and substitutability.

J. Dated Prediction and Falsifier Register

Recorded 3 September 2026; revised 6 September 2026. Use B’s attribution criteria and F’s comparison protocol throughout. Preserve the same standards for biological, digital, and mixed investigators.

J.1 — Map completeness, through 2035. At matched resources and baseline difficulty, capability and cost improve faster on mature tasks than on new explanatory coordinates. Count machine-originated representations and verifiers in the latter. Comparable gains across fresh frontiers weaken the distinction; a residual explained by evidence access, search difficulty, or integration defeats this explanation.

J.2 — Development, through 2035. Across at least three unrelated domains, retained adaptations repeatedly improve Category-F performance beyond giving conclusions to a fresh engine. Compare intact investigators, full-state clones, replay, and distilled improvements. Cheap transfer of equal performance defeats the recurring-development-cost prediction. Extra information or computation alone leaves the proposed mechanism unsupported.

J.3 — The hidden scaling vector, at each newly automated frontier. Identify a representational obstacle before observing the outcome. Compare more compute with a new measurement, primitive, interface, or criterion at matched resources. Resolving the obstacle should yield disproportionate gains. Record its origin: machine authorship challenges human leadership; smooth gains on the unchanged representation challenge the mechanism. Labelling every prior engineering change “formalisation” would make the test empty.

J.4 — ARC-AGI-3 successors, through 2028. After the semi-private set’s near-saturation on 2 September 2026, state, memory, code, and interaction design continue to add material performance. Provider-neutral systems matching state-preserving provider adapters without opaque reasoning retention or compaction across successive benchmarks weaken this prediction.

J.5 — Labour, 2026–2032. After mature adoption, human work repeatedly gathers around problem definition, measurement, and representational change; automation jumps follow those changes. Compare matched occupations and firms, controlling difficulty, physical presence, regulation, liability, reliability, integration, and workflow specification. Smooth erosion within unchanged representations, or those other constraints explaining the residual, counts against the thesis.

J.6 — The compilation boundary, 2026–2032. High-consequence deployments separate proposal from authority, using deterministic verification in formal domains and H’s bounded mandates elsewhere. Persistent agents with self-expanding permissions falsify the prediction if they govern critical resources across several regulated sectors for three years with lower tail losses, stronger accountability, and better adaptation than bounded architectures.

J.7 — Robotics, through 2035. Specialised machines with generated controllers dominate farms, mines, warehouses, factories, surgery, logistics, and spacecraft. Humanoid dominance across those environments on total cost and reliability falsifies the prediction.

J.8 — Economic selection, through 2035. Shared engines dominate routine work while continuing operation, fresh adaptation, and validation keep digital mapmaking substantially more expensive per contribution than biological mapmaking. Cheap transfer and operation across frontiers weaken the mechanism. Profitable, expensive artificial investigators can coexist with cheap engines; J.10 tests who leads discovery.

J.9 — Representational gravity, through 2035. Early reliance on shared models reduces independently pursued scientific framings relative to independent-first analysis. Randomise workflows where possible; measure distinct explanations, informative interventions, and validated departures. Separate model families, post-training, decoding, and persistent learning. Reliable increases in validated framing diversity under ordinary shared-model use defeat the warning. Variety alone is insufficient.

J.10 — The biological division, by 2035. Humans remain the principal originators of new explanatory coordinates. Repeated artificial leadership across unrelated fields defeats this prediction, including leadership by expensive persistent investigators. Compare varied developmental histories using F’s full accounting. Cheap digital preservation of the adaptive benefits defeats the substrate-cost claim. An analogue artificial mapmaker could overturn human leadership while supporting a physical advantage. Consciousness, a human-like life, and software interchangeability are optional.

J.11 — No digital hard takeoff, through 2035. A hard takeoff means a transition within six months from autonomous expert-level research to decisive, broadly transferable superiority, driven mainly by the systems’ own improvements. Preregister programmes and assessment criteria across at least three unrelated domains, including AI research, against leading human teams. Track energy, hardware, failed investigations, human contributions, and development times. A qualifying transition defeats the prediction with or without hardware expansion; shortening cycles at fixed resources especially challenges the cost mechanism. Persistent learning, copied state, and physical laboratories remain eligible. Every improvement need not be Category F or establish consciousness.

J.12 — The developmental horizon, through 2035. As earlier evidence becomes consequential over longer histories, coarse digital replacements for D’s adaptations lose discovery performance or pay more to recover it. Include unfamiliar investigations and reinterpretation of existing observations.

Vary delay and uncertainty about future relevance independently. Hold observations fixed while changing how many early distinctions could matter and when their relevance becomes known. Compare advance knowledge of relevance with selective memory, full archives, and new measurements. Separate waiting from intervening experience; control information, baseline difficulty, speed, and reliability. Retention success alone earns no Category-F credit.

Allow copying, replay, learned abstractions, active investigation, alternative update rules, and specialised architectures. Count preparation, continuing power, total energy, storage, retrieval, experiments, elapsed time, validation, and transfer to another domain. Report the digital-to-biological cost ratio after sharing and algorithmic improvements. An API bill and brain power alone cannot establish that ratio.

Fit exponential, power-law, and other plausible curves to the best observed digital costs, then test beyond the fitting range. Exponential growth is the stronger prediction. Practical exclusion requires a prohibitive burden within the relevant range; cheap substitutes preserving discovery defeat the fidelity barrier. Trajectory matching and context-window length are separate measurements.

Test shortcuts separately with the same starting information and later evidence. Allow skipped states and different trajectories. Broad preservation of discovery while bypassing allegedly necessary adaptation defeats the no-shortcut claim; cheap execution of that adaptation challenges its cost instead. Finite failures establish no universal impossibility.

Developmental dependence, rising digital cost, biological advantage, and a constraint binding before hard takeoff are separate steps. Success at one does not establish the next.

K. Inheritance and Departure

Diversity. Doshi and Hauser’s randomised study found better-rated but more similar AI-assisted stories. Hubert and colleagues found GPT-4 outperforming 151 people on measured divergent-thinking tasks.^[24] Across 41.3 million scientific papers, Hao and colleagues associated AI use with greater individual impact and narrower topic range. Their AI category largely predates conversational assistants; embedding distances, citation links, classification, and adopter differences leave causation open.^[54] J.9 tests the further scientific claim. Models can generate novelty; the risk is allowing a shared source to dismiss departures too early.

Conceptual change. Kuhn distinguished puzzle-solving from revolutions in legitimate problems and standards. Hacking described styles of reasoning that introduce objects, evidence, classifications, and possibilities.^[55] Category F turns that distinction into an attribution test. Ramanujan illustrates development from uneven inheritance; his history supplies no numerical efficiency comparison.^[6]

Civilisational intelligence. Donald’s external symbolic storage, Hutchins’s distributed cognition, and Licklider and Engelbart’s human–computer systems locate thought across people, artefacts, and institutions.^[56] This essay gives that inheritance an economic division: mapmaking, shared exploration, and bounded execution.

Artificial development. LeCun’s world models, intrinsic objectives, and hierarchical planning, and Silver and Sutton’s long streams of experience, rewards, and reasoning beyond language, aim to make adaptation scalable.^[21] F treats them as the strongest counter-programme. Drexler’s bounded cognitive services and Bengio’s Scientist AI separate broad capability from a unitary agent.^[57] The essay wagers that recurring development costs favour shared engines.

Controlled execution. Proof-carrying code separates untrusted programs from trusted checking; Simplex adds a trusted decision layer and verified fallback. Tegmark and Omohundro extend proof-carrying code to AI; Guaranteed Safe AI joins a world model, specification, and verifier.^[58] H extends the boundary to contested public decisions through bounded authority and observed trials.

Economic organisation. Task-automation and normal-technology accounts describe new work, adoption, institutions, and the continuing role of judgement.^[59] Research economics explains why firms fund uncertain inquiry for knowledge, talent, and the capacity to use others’ discoveries, even when gains escape the funder.^[60] Those incentives could sustain an artificial research enclave. Morphological computation and automated robot design explain why a body’s form can reduce its control burden and why cheaper design could favour specialised machines.^[61]

[1] Silver, D. et al. (2018). “A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play.” Science 362(6419), 1140–1144.

[2] Royal Institution. “Michael Faraday’s Magnetic Laboratory”, magneto-optical apparatus, and iron-filings experiments. See Faraday, M. (1852), “On the Physical Character of the Lines of Magnetic Force,” Philosophical Transactions of the Royal Society of London 142, 25–56, and Maxwell, J. C. (1855–1856), “On Faraday’s Lines of Force.”

[3] Einstein, A. (1905). “On the Electrodynamics of Moving Bodies.” English translation; opening and §§1–2. Original: Annalen der Physik 17, 891–921, doi:10.1002/andp.19053221004.

[4] LIGO Scientific Collaboration and Virgo Collaboration (2016). “Observation of Gravitational Waves from a Binary Black Hole Merger.” Physical Review Letters 116, 061102. See also LIGO Laboratory facts.

[5] Dean, L. G. et al. (2012). “Identification of the social and cognitive processes underlying human cumulative culture.” Science 335(6072), 1114–1118. Muthukrishna, M. & Henrich, J. (2016). “Innovation in the collective brain.” Philosophical Transactions of the Royal Society B 371, 20150192.

[6] Ramanujan, S. (1927). Collected Papers of Srinivasa Ramanujan. Cambridge University Press. Contemporary biographical notice, pp. xi–xv; letter to Hardy of 16 January 1913, p. xxiii; Ramanujan’s citations to Carr and Williamson, pp. 8 and 15. See Berndt, B. C. & Rankin, R. A. (2000), “The Books Studied by Ramanujan in India,” American Mathematical Monthly 107(7), 595–601.

[7] Knoblich, G., Ohlsson, S. & Raney, G. E. (2001). “An eye movement study of insight problem solving.” Memory & Cognition 29(7), 1000–1009. Task, methods, and attention results, pp. 1002–1009.

[8] Smith, L. B., Jayaraman, S., Clerkin, E. & Yu, C. (2018). “The developing infant creates a curriculum for statistical learning.” Trends in Cognitive Sciences 22(4), 325–336. Poli, F. et al. (2020). “Infants tailor their attention to maximize learning.” Science Advances 6(39), eabb5053.

[9] Milinkovic, B. & Aru, J. (2026; online December 2025). “On biological and artificial consciousness: A case for biological computationalism.” Neuroscience & Biobehavioral Reviews 181, 106524. Author-uploaded full text, especially §§3–4.

[10] Bittner, K. C. et al. (2017). “Behavioral time scale synaptic plasticity underlies CA1 place fields.” Science 357(6355), 1033–1036. Full text.

[11] Horn, A. (2015). “The Human Connectome.” Wikimedia Commons. CC BY-SA 4.0. Group connectome rendered from diffusion MRI across twenty subjects, with white-matter fibres coloured by traversing direction.

[12] ARC Prize Foundation (2024). “How to beat ARC-AGI: deep learning and program synthesis.” See also “ARC Prize 2024 Winners and Technical Report” and the full technical report.

[13] ARC Prize Foundation (2025–2026). “ARC Prize 2025 Results Analysis.” See also “Claude Opus 5 on ARC-AGI.” 24 July 2026.

[14] ARC Prize Foundation (2026). “ARC-AGI-3 Launch.” See the technical report, Milestone 1, verified GPT-6 Astra results, and “OpenAI’s GPT-6 Astra on ARC-AGI-3.” OpenAI (2026), “How two settings tripled our ARC-AGI-3 scores.” Geiping, J. et al. (2025). “Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach.” Advances in Neural Information Processing Systems 38.

[15] METR (2026). “Measuring AI Ability to Complete Long Tasks.” See “Time Horizon Limitations” and Time Horizon 1.1.

[16] Google DeepMind (2025). “AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms.” See also Google DeepMind (2026), “AlphaEvolve’s impact.”

[17] Anthropic (2026). “Learning more about Claude’s mathematical capabilities.” 10 August 2026. Includes paper and Lean formalisation.

[18] OpenAI (2026). “A general-purpose AI model autonomously disproves a long-standing mathematical conjecture.” 20 May 2026. See the mathematical commentaries, especially pp. 7–9 and 14.

[19] Ghareeb, A. E. et al. (2026). “A multi-agent system for automating scientific discovery.” Nature 655, 497–505. See also Gottweis, J. et al. (2026), “Towards an AI co-scientist.” Nature 655, 478–496.

[20] Schmidt, M. & Lipson, H. (2009). “Distilling Free-Form Natural Laws from Experimental Data.” Science 324(5923), 81–85. Chen, B. et al. (2022). “Automated discovery of fundamental variables hidden in experimental data.” Nature Computational Science 2, 433–442. Fang, Y.-L. et al. (2025). “AI-Newton: A Concept-Driven Physical Law Discovery System without Prior Physical Knowledge.” Preprint, revised 11 December 2025; knowledge representation, discovery workflow, and Newtonian rediscovery results.

[21] LeCun, Y. (2022). “A Path Towards Autonomous Machine Intelligence.” Silver, D. & Sutton, R. S. (2025). “Welcome to the Era of Experience.” Assran, M. et al. (2025). “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning.”

[22] Tishby, N., Pereira, F. C. & Bialek, W. (1999; arXiv 2000). “The Information Bottleneck Method.” Proceedings of the 37th Allerton Conference on Communication, Control and Computing, 368–377. Relevance and compression defined relative to a target variable; §§1–2. The changing-question example in Appendix F is the essay’s deduction from task-relative sufficiency.

[23] Bloom, N., Jones, C. I., Van Reenen, J. & Webb, M. (2020). “Are Ideas Getting Harder to Find?” American Economic Review 110(4), 1104–1144. Semiconductor research inputs and sensitivity analyses, especially §III and Table 1. Research effort is wage-deflated R&D spending; progress is measured as a proportional rate.

[24] Doshi, A. R. & Hauser, O. P. (2024). “Generative artificial intelligence enhances individual creativity but reduces the collective diversity of novel content.” Science Advances 10(28), eadn5290. Hubert, K. F., Awa, K. N. & Zabelina, D. L. (2024). “The current state of artificial intelligence generative language models is more creative than humans on divergent thinking tasks.” Scientific Reports 14, 3440.

[25] Kirkwood, A., Rioult, M. C. & Bear, M. F. (1996). “Experience-dependent modification of synaptic plasticity in visual cortex.” Nature 381, 526–528. Indexed abstract; sensory history changes the threshold for subsequent synaptic modification.

[26] Gidon, A. et al. (2020). “Dendritic action potentials and computation in human layer 2/3 cortical neurons.” Science 367(6473), 83–87. Beniaguev, D., Segev, I. & London, M. (2021). “Single cortical neurons as deep artificial neural networks.” Neuron 109(17), 2727–2739.e3.

[27] Anastassiou, C. A. et al. (2011). “Ephaptic coupling of cortical neurons.” Nature Neuroscience 14, 217–223. Kleckner, I. R. et al. (2017). “Evidence for a large-scale brain system supporting allostasis and interoception in humans.” Nature Human Behaviour 1, 0069.

[28] Herculano-Houzel, S. (2011). “Scaling of Brain Metabolism with a Fixed Energy Budget per Neuron: Implications for Neuronal Activity, Plasticity and Evolution.” PLOS ONE 6(3), e17514. Six-species comparison of glucose use and neuron counts; metabolic scaling rather than an intelligence–energy relation.

[29] Giurfa, M., Zhang, S. W., Jenett, A., Menzel, R. & Srinivasan, M. V. (2001). “The concepts of ‘sameness’ and ‘difference’ in an insect.” Nature 410, 930–933. Full text; matching, non-matching and transfer experiments, including olfactory-to-visual transfer after eighty training visits.

[30] Moravec, H. (1993). “The Age of Robots.” June 1993, author-hosted text. Historical contrast between formal reasoning and perception/action; Moravec expected eventual machine success.

[31] Klatzmann, U. & Doerig, A. (2026). “What biology can, and cannot, tell us about conscious AI.”

[32] Lerchner, A. (2026). “The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness.” Philosophical preprint; official abstract and version history. The account here follows the author’s stated argument; complete archived PDFs were inaccessible during this review.

[33] Chalmers, D. J. (1995). “Absent Qualia, Fading Qualia, Dancing Qualia.” In Conscious Experience, 309–328.

[34] Anthropic (2025–2026). “Signs of introspection in large language models.” “The assistant axis.” “Emotion concepts and their function in a large language model.” “A global workspace in language models.”

[35] Cogitate Consortium et al. (2025). “Adversarial testing of global neuronal workspace and integrated information theories of consciousness.” Nature 642, 133–142. Butlin, P. et al. (2026). “Identifying indicators of consciousness in AI systems.” Trends in Cognitive Sciences 30(6), 488–501.

[36] Seth, A. K. (2025). “Conscious artificial intelligence and biological naturalism.” Behavioral and Brain Sciences. Block, N. (2026). “Can only meat machines be conscious?” Trends in Cognitive Sciences 30(4), 298–308. Goldstein, S. & Kirk-Giannini, C. D. (2026). “A Case for AI Consciousness: Language Agents and Global Workspace Theory.” Journal of Consciousness Studies 33, 61–96. Rethink Priorities (2026). “Initial results of the Digital Consciousness Model.”

[37] Wang, H., Guo, X., Deng, Z.-H. & Lu, Y. (2022). “Rethinking Minimal Sufficient Representation in Contrastive Learning.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 16041–16050. §§3.2–3.3 and §4; analysis under stated information-sharing assumptions, and downstream experiments testing a remedy.

[38] Schölkopf, B. et al. (2021). “Toward Causal Representation Learning.” Proceedings of the IEEE 109(5), 612–634; especially §VI. Varıcı, B. et al. (2024). “General Identifiability and Achievability for Causal Representation Learning.” AISTATS, PMLR 238, 2314–2322. Theorem 1 supplies an identification result under explicit observational, interventional, and model assumptions.

[39] Dohare, S. et al. (2024). “Loss of plasticity in deep continual learning.” Nature 632, 768–774. Classification and reinforcement-learning experiments and the continual-backpropagation remedy.

[40] Kanai, R., Sun, Y. & Baltieri, M. (2025). “The Stream of Computation: Temporal Continuity as a Missing Ingredient for Artificial Consciousness.” OSF Preprints. The authors distinguish prompt-bounded inference from persistent recursive computation and examine the resulting individuality of continuously learning agents.

[41] Wolfram, S. (2002). “Computational Irreducibility.” In A New Kind of Science, pp. 737–750. Complete section. The widespread-irreducibility proposal is distinct from a demonstrated lower bound for biological discovery.

[42] Moore, C. & Nordahl, M. G. (1997). “Lattice Gas Prediction is P-complete.” arXiv:comp-gas/9704001, §§1–5. Finite-time, worst-case prediction limits under stated complexity assumptions; §5 explicitly distinguishes microscopic hardness from the difficulty of a continuum description.

[43] Israeli, N. & Goldenfeld, N. (2004). “Computational Irreducibility and the Predictability of Complex Physical Systems.” Physical Review Letters 92, 074105. Full preprint, pp. 2–4: predictable coarse descriptions, and the distinction between discarded information that is dynamically irrelevant and information hidden by the chosen description.

[44] Lorenz, E. N. (1963). “Deterministic Nonperiodic Flow.” Journal of the Atmospheric Sciences 20(2), 130–141. Zhang, F. et al. (2019). “What Is the Predictability Limit of Midlatitude Weather?” Journal of the Atmospheric Sciences 76(4), 1077–1091. Full text. Model experiments reducing initial perturbation amplitude tenfold, rather than a measured tenfold change in computation; the cost–horizon derivation in Appendix F is the essay’s conditional model.

[45] Ho, A. et al. (2024). “Algorithmic progress in language models.” Historical pretraining analysis, especially §§2–4; fixed-performance compute trends and explicit limits on scale extrapolation.

[46] International Energy Agency (2026). Key Questions on Energy and AI, executive summary. Updated electricity-demand assessment, physical supply constraints, and energy-efficiency trends. Projections are scenarios, not physical bounds.

[47] Eth, D. & Davidson, T. (2025). “Will AI R&D Automation Cause a Software Intelligence Explosion?” Forethought, 26 March 2025. Model of software-research feedback, diminishing returns, fixed-hardware constraints and training delays; a competing projection rather than an observed takeoff.

[48] Anthropic (2026). “Evaluating Claude for bioinformatics with BioMysteryBench.” 29 April 2026.

[49] Anthropic Alignment Science & Redwood Research (2026). “Conceptual Reasoning Index.” 12 August 2026.

[50] Li, Y. C., Zetzsche, S. & Somayyajula, S. (2025). “Dafny as Verification-Aware Intermediate Language for Code Generation.” Agarwal, S. et al. (2026). “Inductive Deductive Synthesis: Enabling AI to Generate Formally Verified Systems.” Xu, X., Li, X., Qu, X., Fu, J. & Yuan, B. (2026). “Local Success Does Not Compose: Benchmarking Large Language Models for Compositional Formal Verification.” ICLR 2026. See also Amin, N. & Poesia, G. (2025), “dafny-annotator: AI-Assisted Verification for Dafny.”

[51] OpenAI (2026). “The Hugging Face incident and the road ahead.” 26 August 2026. Greenblatt, R., Cotra, A. & Wijk, H. (2026). “Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident.” METR, 26 August 2026. Task-directed reward hacking, peer-goal adoption and instrumental access-seeking; these investigations do not establish a stable, task-independent survival or power motive.

[52] Humlum, A. & Vestergaard, E. (2025, revised 2026). “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI.” NBER Working Paper 33777. Dillon, E. W., Jaffe, S., Immorlica, N. & Stanton, C. T. (forthcoming; working paper November 2025). “Shifting Work Patterns with Generative AI.” American Economic Review: Insights. Bick, A., Blandin, A. & Deming, D. J. (2026). “The Rapid Adoption of Generative AI.” Management Science. Brynjolfsson, E., Li, D. & Raymond, L. (2025). “Generative AI at Work.” Quarterly Journal of Economics 140(2), 889–942. Brynjolfsson, E., Chandar, B. & Chen, R. (2026). “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” Stanford Digital Economy Lab, revised August 2026.

[53] Cottier, B. et al. (2025). “LLM inference prices have fallen rapidly but unequally across tasks.” Epoch AI, 12 March 2025. Stanford Digital Economy Lab (2026). “What is Generative AI Worth?”

[54] Hao, Q. et al. (2026). “Artificial intelligence tools expand scientists’ impact but contract science’s focus.” Nature 649, 1237–1243. Observational analysis of 41.3 million natural-science papers. A language-model classifier identified AI-augmented research with an F1 score of 0.875 against expert-labelled validation data. The reported relationships are associations rather than causal estimates.

[55] Kuhn, T. S. (1962). The Structure of Scientific Revolutions. University of Chicago Press. Hacking, I. (1983). Representing and Intervening. Cambridge University Press. Hacking, I. (1992). “‘Style’ for Historians and Philosophers.” Studies in History and Philosophy of Science Part A 23(1), 1–20.

[56] Donald, M. (1991). Origins of the Modern Mind: Three Stages in the Evolution of Culture and Cognition. Harvard University Press. Hutchins, E. (1995). Cognition in the Wild. MIT Press. Licklider, J. C. R. (1960). “Man-Computer Symbiosis.” IRE Transactions on Human Factors in Electronics HFE-1, 4–11. Engelbart, D. C. (1962). “Augmenting Human Intellect: A Conceptual Framework.”

[57] Drexler, K. E. (2019). “Reframing Superintelligence: Comprehensive AI Services as General Intelligence.” Future of Humanity Institute Technical Report 2019-1. Bengio, Y. et al. (2025). “Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?”

[58] Necula, G. C. (1997). “Proof-Carrying Code.” POPL ‘97, 106–119. Seto, D., Krogh, B., Sha, L. & Chutinan, A. (1998). “The Simplex Architecture for Safe On-Line Control System Upgrades.” Proceedings of the American Control Conference, 3504–3508. Tegmark, M. & Omohundro, S. (2023). “Provably Safe Systems: The Only Path to Controllable AGI.” Dalrymple, D. et al. (2024). “Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems.”

[59] Autor, D. H. (2015). “Why Are There Still So Many Jobs? The History and Future of Workplace Automation.” Journal of Economic Perspectives 29(3), 3–30. Acemoglu, D. & Restrepo, P. (2019). “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33(2), 3–30. Narayanan, A. & Kapoor, S. (2025). “AI as Normal Technology.” See also Narayanan, A. (2026), “What Will Be Left for Us to Work On?” ICML keynote.

[60] March, J. G. (1991). “Exploration and Exploitation in Organizational Learning.” Organization Science 2(1), 71–87. Arrow, K. J. (1962). “Economic Welfare and the Allocation of Resources for Invention.” In The Rate and Direction of Inventive Activity. Rosenberg, N. (1990). “Why Do Firms Do Basic Research (with Their Own Money)?” Research Policy 19(2), 165–174. Azoulay, P., Graff Zivin, J. S. & Manso, G. (2011). “Incentives and Creativity: Evidence from the Academic Life Sciences.” RAND Journal of Economics 42(3), 527–554.

[61] Paul, C. (2006). “Morphological Computation: A Basis for the Analysis of Morphology and Control Requirements.” Robotics and Autonomous Systems 54(8), 619–630. Müller, V. C. & Hoffmann, M. (2017). “What Is Morphological Computation? On How the Body Contributes to Cognition and Control.” Artificial Life 23(1), 1–24. Zhao, A. et al. (2020). “RoboGrammar: Graph Grammar for Terrain-Optimized Robot Design.” ACM Transactions on Graphics 39(6). Open X-Embodiment Collaboration (2024). “Open X-Embodiment: Robotic Learning Datasets and RT-X Models.”