Projected Impact of Superhuman AI: A Study of Capability, Diffusion and Governance

An evidence-based analysis of heterogeneous machine capabilities, economic transmission and conditions for accountable research acceleration

Abstract

What would have to be true for increasingly capable artificial intelligence to produce an economic transformation comparable to earlier industrial revolutions? This study decomposes that proposition into measurable claims about task performance, dependable autonomy, organisational adoption, research productivity and institutional control. Scaling experiments, instruction-following models, code generation and algorithm discovery establish important capabilities in specified settings. They do not by themselves establish broad occupational substitution or self-sustaining research acceleration. Evidence from machine translation, industrial robotics and AI-related hiring illustrates distinct channels through which automation can affect economic activity, while leaving the consequences of general-purpose language models substantially unresolved. Economic effects depend on complementary investment, demand, bargaining institutions and the distribution of adoption. A proposed prospective programme evaluates quality-adjusted outputs, complete costs, research validity and worker outcomes while maintaining a versioned forecast register. Governance is treated as a determinant of the development path, including accountability for delegation, documentation, evaluation and recovery. The central conclusion is that transformative impact is a conditional system-level proposition. Its strongest tests concern the conversion of technical capability into verified, affordable and widely usable work, rather than a single date on which machines become categorically superhuman.

Contents

Research orientation

The study begins with measurable definitions and the available technical evidence, then examines economic applications, diffusion and research feedback. The final sections propose tests and institutional arrangements for accountable deployment.

1 Research question and scope

A forecast of superhuman AI contains several propositions that can come apart. Machines might outperform specialists on selected tasks without replacing their occupations. Firms might automate substantial work without generating an immediate increase in aggregate wages. AI might contribute original algorithms without controlling the research process that produces its successors. The organising question is therefore not simply when a model becomes more intelligent than a person. It is which capabilities, organisational changes and institutional arrangements would jointly produce a sustained transformation in the production and distribution of useful knowledge.

This study is the analytical companion to LupoToro’s opinion essay on the projected impact of superhuman AI. Its purpose is to turn the broader argument into distinct, testable propositions. Technical results, economic observations and theoretical models are assessed according to the questions each can answer. No proprietary forecasting model, internal experiment, strategic simulation or record of successful company predictions is assumed.

The term transformation requires an outcome and a comparison period. It could refer to a higher level of output, a permanently faster growth rate, a redistribution of income, a change in the organisation of research, or a concentration of decision-making authority. These are different outcomes. A one-off reduction in the cost of producing a document is not the same as an enduring increase in the rate of scientific discovery. Comparing AI with the Industrial Revolution without distinguishing these quantities makes the comparison rhetorically powerful but empirically indeterminate.

The working thesis is that the consequential transition occurs when reliable capabilities become affordable components of institutions that can use, verify and contest their outputs. This allows rapid technical change and slow organisational change to coexist. It also explains why an important technology can create both local productivity gains and substantial adjustment costs. The analysis assesses these mechanisms without assigning unsupported dates, national rankings or probabilities to a future general-purpose intelligence.

2 Evidence selection and analytical method

This is a structured critical synthesis with a proposed prospective research programme. It does not report a completed LupoToro experiment, a systematic review of every AI publication or a pooled causal estimate. Source selection follows the causal chain from model development to task performance, workplace adoption, research output and governance. Original experiments, theoretical contributions, research-institution working papers and official technical frameworks are prioritised. Publication records, reported outcomes and accessible methods were checked; findings are confined to what those records establish.

The available research provides reasons to expect further progress and reasons to reject a simple parameter-count forecast. Kaplan et al. (2020) documented empirical relationships between language-model cross-entropy loss, model size, data and training compute. Hoffmann et al. (2022) showed that the allocation of a given compute budget between parameters and training data materially changed performance. These results concern particular model families, training regimes and endpoints. They do not supply a law converting a specified quantity of compute into an occupational capability or a date for general intelligence.

Prediction performance and useful behaviour are different objectives. Ouyang et al. (2022) reported that human-feedback training improved preference ratings and selected behavioural measures, while the resulting systems continued to make mistakes. Such findings support targeted improvements in instruction following. They do not establish an enduring commitment to a developer’s intentions across every future context, nor do they allow a system’s motives to be read directly from fluent explanations.

Evidence is evaluated according to design and scope. A benchmark tests the evaluated task distribution and tool configuration. A workplace intervention can support a causal interpretation only under its identification assumptions; comparisons between more and less exposed regions or establishments require attention to confounding and counterfactual trends. A growth model yields implications conditional on its structure. None of these designs alone establishes the complete path from technical improvement to national productivity. Preprints and working papers are identified by their publication status, and proposed experiments are distinguished from completed studies.

3 Defining superhuman capability and dependable autonomy

Superhuman should describe a comparison, not a substance possessed by a machine. A meaningful claim names the task distribution, human comparator, quality threshold, available information, tools, time allowance and cost. Beating an average unassisted participant differs from beating an experienced professional using appropriate tools. Exceeding the best observed score on one task differs from superior performance across a representative portfolio. Relevant comparisons should therefore include unaided humans, humans with established software and humans using the candidate AI system.

Performance is better represented by a profile than a single rank. Accuracy, coverage of tasks, resource use and the amount of human assistance answer different questions. A system may perform well when its inputs and objective are tightly specified yet fail when it must identify what problem to solve. This study therefore uses task breadth, achieved quality, required resources and permitted intervention as separate descriptive dimensions. They are a proposed evaluation scheme, not a claim that an agreed universal intelligence scale has been established.

AlphaGo Zero demonstrates that superhuman performance can be established within a demanding but well-defined domain. It learned through self-play and surpassed earlier champion-level Go systems (Silver et al., 2017). The game supplies fixed rules, observable outcomes and repeated opportunities to evaluate improvement. That achievement is substantial on its own terms. Transferring its conclusion to professional work would require evidence about tasks with incomplete specifications, delayed feedback and consequences beyond winning a game.

Benchmarks remain useful when their interpretation is specific. The MMLU benchmark introduced by Hendrycks et al. (2021) tests knowledge and problem-solving across 57 subjects. Success on such questions can establish breadth within the tested format. It cannot alone establish that a system can manage a laboratory, negotiate ambiguous requirements or assume responsibility for a professional service. Fresh tasks, contamination checks, appropriate human baselines and independent scoring are needed when making a broader claim.

Autonomy adds a temporal and organisational dimension. A system that produces a useful suggestion transfers verification and execution to someone else. A delegated agent must also preserve constraints, recognise missing information, track commitments and recover appropriately. Its performance should be measured as completed work with the required quality, including the cost of interventions and failures. A collection of individually impressive actions may still form an unreliable process if errors propagate or important exceptions are missed.

Capability and authority should be specified separately. Giving a model access to more tools can improve task completion while increasing the consequences of a mistake. A deployment decision therefore cannot be derived solely from the model’s rank. Equally, claims about agency, deception or consciousness require their own operational definitions and evidence. Successful execution does not establish subjective experience, and a failure to follow instructions does not, without further evidence, establish a stable adversarial objective.

4 Evidence from technical and economic applications

The available evidence spans different units of analysis: generated programs, competitive problem solving, transactions on a digital platform and employment in local labour markets. These outcomes cannot be combined into a universal productivity multiplier. Their value is to identify which links in the causal chain have been tested and which remain prospective. An impressive technical result may justify a workplace experiment without already establishing its result.

Chen et al. (2021) evaluated language models trained on code using HumanEval, a set of programming tasks assessed for functional correctness. Repeated sampling improved the chance of finding a passing solution, while the authors also identified failures on complex specifications. This is evidence about program synthesis under a defined testing procedure. A success rate allowing many candidate programs is a different quantity from the reliability of one unchecked answer, and neither directly measures the productivity of a developer maintaining an unfamiliar software system.

AlphaCode extends this research to competitive programming. Li et al. (2022) report an average placement in the top 54.3% of competitors in simulated Codeforces evaluations. The system generates a large candidate pool and filters and clusters it into a small set of submissions. This establishes competitive performance in the evaluated setting, not superiority to the best programmers. Commercial software work additionally involves requirements, integration, maintenance and coordination; those outcomes are not measured by a programming-contest ranking.

There is also evidence that a bounded AI improvement can affect actual transactions. Brynjolfsson, Hui and Liu (2019) study the introduction of improved machine translation on a large digital trading platform and find increased cross-border trade. Their identification strategy exploits the platform change to examine a specific language barrier. The study supports an economic channel through reduced communication costs. It does not measure general-purpose cognitive substitution, national productivity or the employment effects of conversational language models.

Industrial robotics provides a different comparison. Acemoglu and Restrepo (2020) examine exposure to robot adoption across US local labour markets and estimate adverse employment and wage effects in more exposed areas. This is evidence that automation’s productivity potential need not translate into favourable outcomes for every affected worker. Industrial robots, the period studied and the local labour-market design are integral to the result. Applying its coefficients to language models or future research systems would require an additional model and evidence.

The studies support separating technical feasibility, practical integration and economic incidence. They do not resolve the magnitude of a general-purpose language model’s contribution to professional work. That question calls for direct comparisons of completed work, including review and recovery, with appropriate controls and representative users. A reduction in time spent drafting code or text is one possible mechanism to investigate; a measured reduction in the complete cost of a service is a stronger claim.

One prospective hypothesis is that assistance will be most useful when requirements are clear and errors are inexpensive to detect. Another is that integration costs will dominate where context is extensive or mistakes are costly. These predictions should be tested within a common design rather than inferred by comparing unrelated studies. Tasks, users and the evaluation period must be defined before observing results. Both benefits and failures should remain part of the evidence used to assess diffusion.

5 From technical capability to economic diffusion

Economic exposure, adoption and displacement are separate variables. An occupation contains multiple tasks, including responsibilities that are difficult to observe in standard descriptions. Felten, Raj and Seamans (2021) construct exposure measures linking AI applications to occupational abilities and extend them to industries and geographic areas. Their measure is agnostic about whether an affected ability is complemented or substituted. Exposure therefore identifies where change may be feasible; it does not establish jobs lost or the speed with which employers will reorganise work.

The task framework of Acemoglu and Restrepo (2019) distinguishes the displacement of labour from existing tasks, productivity effects and the creation of new tasks in which labour is employed. These mechanisms can operate simultaneously. Cheaper production may expand demand; automation may also reduce labour’s share of a given activity. Neither universal unemployment nor automatic compensation of every displaced worker follows from a reduction in task costs. Distributional outcomes require evidence about prices, demand, new activities and bargaining arrangements.

Complementary investment creates another delay. Brynjolfsson, Rock and Syverson (2021) model how investment in organisational and other intangible assets can precede the measured productivity benefits of a general-purpose technology. The implication for AI is a testable mechanism, not a licence to explain away every disappointing result as a temporary lag. A firm claiming delayed benefits should identify the investment, its expected transmission and the outcomes that would show whether the reorganisation worked.

Acemoglu et al. (2022) examine US online vacancies over 2010–2018. AI-exposed establishments show growing demand for AI skills alongside reduced hiring in non-AI positions and changing skill requirements. The authors do not detect corresponding aggregate employment or wage effects in more exposed occupations and industries. This evidence concerns the technologies and adoption observed in that period. It cannot settle the consequences of current language models, but it illustrates why changes in establishments and aggregate outcomes should be measured separately.

A useful diffusion model therefore separates technical feasibility from commercial implementation. An employer needs access to usable systems, relevant data, integration capacity, management attention and a way to handle failure. Workers need time and incentives to learn. Customers may value accountability, confidentiality or continuity even when an automated output is cheaper. Adoption can consequently be rapid at the level of individual experimentation and slower at the level of redesigning an entire service.

Employment effects also depend on the product market. For a fixed product mix, total labour hours equal output multiplied by labour hours per unit. An improvement that reduces the second term can coincide with rising employment if demand expands enough to increase the first. How much cost reduction reaches customers depends on competition and pricing; how much benefit reaches workers depends on remuneration and bargaining. This accounting relationship supplies no numerical forecast. It identifies quantities that a displacement study must observe, including changes among competitors and entrants rather than only the adopting firm.

The distribution of infrastructure and information also matters. Jones and Tonetti’s (2020) theoretical analysis shows how nonrival data can create social gains from reuse while incentives to withhold it can restrict those gains. Applied to AI, this suggests examining access and control alongside technical capability. It does not imply that all data should be shared: privacy, consent and other interests remain part of the allocation problem. Nor does inexpensive copying of information make computing resources, organisational expertise or physical infrastructure costless.

A further research question concerns entry into expertise. If assistants improve novice performance, they might widen access to skilled work. If organisations remove the tasks through which novices learn, they might weaken the future supply of experts able to review difficult cases. Both mechanisms are plausible and can coexist. Longitudinal studies should therefore measure progression, independent competence and opportunities to exercise judgement, alongside immediate output. A short-run gain in throughput is not sufficient evidence about the reproduction of professional capability.

6 AI research acceleration and limits to feedback

The strongest argument for unusually rapid change is that AI can improve the process that produces further AI. Yet this feedback is not one indivisible mechanism. It includes reading, proposing hypotheses, writing software, allocating experiments, interpreting results, selecting research directions and implementing discoveries. These activities have different verification costs and dependencies. Faster coding can increase experimental throughput without proportionately increasing the number of important, reproducible findings.

Machine contributions to algorithm discovery provide a concrete basis for investigating research acceleration. Fawzi et al. (2022) used reinforcement learning in AlphaTensor to discover provably correct matrix-multiplication algorithms, including improvements for particular matrix sizes and arithmetic settings. This is evidence of original computational discovery within a formally specified search space. It does not demonstrate autonomous control of general scientific inquiry or a universal improvement in the asymptotic complexity of matrix multiplication.

A second example concerns prediction in a scientific field. Jumper et al. (2021) report AlphaFold’s performance on the blind CASP14 protein-structure assessment, with accuracy competitive with experimental structures in a majority of evaluated cases. The result demonstrates a major advance in a defined scientific task. It does not establish that structural prediction replaces experiments on function, resolves every biological question or automates the organisation of research. Its broader economic contribution will depend on how reliably predictions support subsequent scientific work.

Algorithmic progress must also be distinguished from additional spending. Hernandez and Brown (2020) measured reductions in training compute needed to reach a fixed historical ImageNet performance target, reporting a 44-fold reduction between 2012 and 2019. This provides an operational example of algorithmic efficiency. The measured trajectory is not a universal doubling law for intelligence, and a gain at a fixed target should not be conflated with an equal improvement at the frontier of every task.

Research returns can change as knowledge accumulates. Bloom et al. (2020) documented declining research productivity across several historical settings even as research effort increased. Those observations do not prove that AI cannot alter the trend. They show why increasing the number or speed of research attempts cannot be assumed to translate linearly into discoveries. Measuring automated research requires a denominator that includes failed experiments, verification, compute and human supervision, as well as successful demonstrations.

Aghion, Jones and Jones (2017) analysed growth models in which automation affects both production and the creation of ideas. Their theoretical cases make bottlenecks and substitution assumptions central to the resulting growth path. This is the appropriate status of an explosive-growth claim: a conditional implication of a specified model, not an observed consequence of every advance in AI. If essential experiments, capital construction or decision processes remain slow, rapid progress in another component may leave the overall cycle constrained.

A proposed empirical test would compare independent research teams assigned similar, previously unresolved problems and equivalent resource budgets. The primary outcome would be independently reproduced progress against preregistered criteria, assessed without revealing the treatment. Counting generated papers, code or plausible hypotheses would invite metric gaming. Follow-up would determine whether improvements transfer to new problems and survive implementation. Acceleration of the complete research cycle is established only when these downstream checks are included; a faster proposal stage remains valuable, but is a narrower result.

7 A falsifiable programme for forecasting impact

A useful forecast is a collection of claims with resolution rules. The proposed programme would maintain a register identifying each outcome, population, measurement method, deadline, assumptions and conditions under which the claim would be judged false. Forecasts about capability, adoption, economic outcomes and governance would be recorded separately. This prevents a prediction of technical progress from being retrospectively treated as confirmation of a different prediction about employment or political control.

The first hypothesis concerns dependable task substitution: a defined class of systems can complete a representative portfolio of work at a lower total cost than an appropriate human baseline, while satisfying the same quality and risk constraints. Tasks would be sampled before treatment assignment, with an independent reserve for testing generalisation. The protocol would count clarification, tool failures, retries, human rescue and final review. Its results would report a capability profile and uncertainty, including failure modes that disappear in a single average.

The second hypothesis concerns organisational complementarity: structured integration produces more useful output than merely providing access to a tool. A factorial design could vary tool access and a clearly specified integration intervention, such as training and workflow redesign. Randomisation at team or workplace level may be needed to limit spillovers. Outcomes would include independently assessed output, costs, employee learning and customer experience. Intent-to-treat estimates and effects among adopters would be distinguished, with the assumptions behind the latter made explicit.

The third hypothesis concerns research feedback: AI assistance increases validated research progress per resource across repeated cycles. Evaluation would include problem selection, execution, verification and deployment rather than only code generation. Teams should not choose their success metric after seeing a promising result. A replication group would test transfer to held-out settings; an economic assessment would include the cost of the evaluator and the preparation of the research environment. Gains dependent on extensive uncounted expert setup would be reported as such.

Forecast assessment needs both an archive and a revision policy. The original wording and evidence set would remain inspectable. New evidence could justify a revised forecast, but the revision would receive its own timestamp and would not replace the original for scoring. Where probabilities are elicited, calibration should be evaluated across many resolvable events rather than celebrated after one correct outcome. This study supplies no numerical probability distribution because none has been estimated or elicited through such a process.

The programme should also distinguish uncertainty from disagreement about values. Whether a system meets a measured error threshold is an empirical question; which consequences society is willing to accept is partly normative. Publishing the measurement does not settle the policy choice. Conversely, disagreement about acceptable risk should not justify changing the measurement or selecting an easier comparator. Separating these roles makes both scientific revision and democratic contestation more coherent.

8 Governance, institutional control and defence-adjacent assurance

A capability forecast is incomplete if institutions are treated as passive recipients of technology. Access rules, procurement, liability, research funding, competition and worker participation can change the development path. Governance is therefore both a response to capability and a causal input into which capabilities are developed and deployed. Analysis should examine these mechanisms rather than treating a competitive race between named states or companies as an inevitable sequence of events.

A useful technical starting point is the distinction between an objective and the consequences of optimising it. Amodei et al. (2016) identify research problems involving unwanted side effects, reward misspecification, costly supervision, unsafe exploration and distributional change. These are mechanisms through which a system can fail without requiring an attribution of conscious intent. Their relevance is prospective: evaluation should examine whether the specified objective, supervision and operating conditions adequately represent the intended task. The paper is a research agenda, not a certificate that these problems have been solved.

Institutional design needs a corresponding record of decisions. Raji et al. (2020) propose internal algorithmic auditing across the development lifecycle, with documentation that links decisions and responsibilities. Mitchell et al. (2019) propose model cards describing intended uses, evaluation conditions and relevant performance differences. These frameworks make limitations and accountability more inspectable; they do not demonstrate that any particular deployment is safe. A proposed assurance process should test whether documentation changes decisions and enables correction, rather than treating the existence of a report as the outcome.

For defence-adjacent research, appropriate applications include logistics, maintenance analysis, administrative support and the organisation of scientific information. Procurement should specify the allowed function, data provenance, review burden and recovery process. An impressive general benchmark cannot by itself establish suitability for a consequential decision. This study makes no operational recommendation about weapons, offensive cyber activity or biological capabilities; the relevant research contribution is the evaluation of reliability, controllability and accountable human judgement.

Concentration also raises a control problem independent of model behaviour. If an institution cannot inspect performance, change suppliers, retain necessary records or operate during a service interruption, it may become dependent even on a well-behaved system. Proposed safeguards should therefore be evaluated by their observable effects: whether independent assessment is possible, whether responsibility is assigned, whether affected people can challenge consequential outputs and whether exit is technically and organisationally feasible. These are testable institutional properties, not assurances delivered by a policy statement alone.

International uncertainty should motivate robust arrangements rather than invented forecasts of espionage or conflict. Shared incident categories, comparable evaluations and protected channels for reporting failures could support cooperation without requiring agreement on every long-term prediction. The feasibility and consequences of each arrangement require separate research. Neither automatic convergence towards common rules nor unavoidable escalation follows from technical progress itself.

9 Infrastructure, aerospace and human performance

Scaling a capable model into a dependable service requires more than replicating its weights. Computing, storage, network access, power, cooling, maintenance and staff are real inputs. Training and inference create different demands; a cheaper training run need not imply a cheaper service under a particular workload. Investment analysis should compare useful output at specified quality and latency, including utilisation and the cost of redundancy. A claim about virtual labour supply cannot omit the physical and organisational resources needed to sustain it.

Aerospace provides a useful test of the distinction between advice and delegated authority. A model that assists documentation or analyses maintenance records performs a different function from a component that selects or executes a consequential action. Parasuraman, Sheridan and Wickens (2000) distinguish information acquisition, analysis, decision selection and action implementation, and explain that automation changes human activity and coordination demands. This framework supports function-specific assessment rather than a binary choice between manual and autonomous systems.

A proposed human-factors study would measure whether operators correctly detect unsuitable recommendations, retain an accurate understanding of system state and recover from a degraded mode. Faster responses would be analysed alongside decision quality and workload. The evaluation should include plausible exceptions and interruptions, since a quiet nominal demonstration may reveal little about recovery. A request for human approval is not an adequate safeguard if the person lacks the information, time or expertise needed to review it.

The organisation must preserve the ability to operate when automation is unavailable. That may require training, documentation and exercises with alternative pathways. These activities consume resources and should appear in the economic comparison. They may also create value beyond measured average productivity by reducing exposure to rare disruptions. A deployment can therefore be worthwhile without having the highest normal-condition throughput, provided that the reliability benefit is specified and supported.

These considerations connect technical progress to professional development. Effective oversight is a skill that requires practice and feedback; it cannot simply be assigned to a worker after the work needed to develop it has disappeared. Research should track whether teams become better at assessing evidence or merely faster at accepting suggestions. The question is not whether people or machines deserve primacy in the abstract, but which allocation of functions produces competent performance, intelligible responsibility and recoverable failure in the intended environment.

10 Interpretation and research priorities

The evidence establishes strong performance in selected computational tasks, original algorithmic discovery and identifiable economic effects of particular AI and automation applications. It provides a basis for investigating wider transformation while leaving major links unresolved. A technology can be powerful without producing a uniform effect across occupations, firms or workers. Treating superhuman performance as a single attribute obscures the variables needed to predict where it will matter.

The most consequential uncertainty concerns the interaction of three transitions. First, capabilities must become sufficiently reliable across useful tasks. Second, organisations must turn them into productive processes rather than additional streams of unverified output. Third, gains must diffuse through institutions that determine access, remuneration and authority. Progress in one transition can accelerate another, but can also create new bottlenecks. The empirical priority is to measure these interactions with explicit counterfactuals.

The strongest rapid-transformation hypothesis is that AI improves both production and the research process, with falling costs and expanding access reinforcing one another. The strongest constraint is that verification, physical implementation and organisational change remain essential and difficult to accelerate. Neither should be accepted by assertion. Prospective studies can determine where the feedback is measurable, where it saturates and whether its gains survive outside unusually favourable settings.

For LupoToro’s analytical agenda, the appropriate output is a conditional forecast map with measurable resolution criteria. Each proposition should name the enabling conditions, leading indicators and evidence that would weaken it. Employment forecasts should distinguish task change, headcount, wages and entry opportunities. Research forecasts should distinguish proposal volume from validated progress. Governance forecasts should identify institutional choices instead of treating political reactions as consequences already determined by model capability.

Transformative AI is therefore a serious research proposition whose assessment requires technical measurement and political economy together. A useful study does not need to settle the arrival date of a universal intelligence. It needs to show how claims about capability connect to observable consequences, where the evidence currently stops and which next observations would change the conclusion. That discipline preserves ambitious inquiry while making both favourable and adverse scenarios open to correction.

Appendix A Measurement and inference audit

The following identities are analytical checks, not estimates from an internal simulation. Their purpose is to reveal assumptions that a forecast or productivity comparison must make explicit.

For a fixed volume of work meeting an agreed quality threshold, let total cost include model access or compute, human preparation, review, rework and allocated integration costs. The relevant comparison is total cost per accepted unit, not cost per generated unit. If an assistant reduces drafting time but increases verification time, both changes enter the same ledger. Costs of failed attempts cannot be omitted merely because they produced no accepted output.

For a simple chain requiring every step to succeed, with independent steps each succeeding with probability p, the probability of complete success is p to the power n, where n is the number of steps. The illustrative values p = 0.99 and n = 100 yield approximately 0.366. This is not an estimate for any AI system. Real workflows include correlated errors, retries, checks and alternative paths; direct measurement is needed. The example shows why a local success rate does not automatically establish long-horizon reliability.

For a serial research process, if fraction f of baseline time is accelerated by factor s and all other costs stay unchanged, total speedup is 1 / [(1 − f) + f/s]. This accounting bound does not model scientific discovery itself. It shows why accelerating one stage cannot be equated with accelerating the complete cycle. A forecast of compounding research progress must additionally specify discovery quality, transfer, verification and resource constraints.

Finally, distinguish a level effect from a growth-rate effect. Producing the same useful output with fewer resources once can raise productivity without making its rate of improvement permanently faster. A claim of sustained acceleration requires repeated observations and a credible account of the counterfactual trend. A single successful demonstration cannot identify that trend.

References

Acemoglu, D. and Restrepo, P. (2019) ‘Automation and new tasks: How technology displaces and reinstates labor’, Journal of Economic Perspectives, 33(2), pp. 3–30. https://doi.org/10.1257/jep.33.2.3

Acemoglu, D. and Restrepo, P. (2020) ‘Robots and jobs: Evidence from US labor markets’, Journal of Political Economy, 128(6), pp. 2188–2244. https://doi.org/10.1086/705716

Acemoglu, D., Autor, D., Hazell, J. and Restrepo, P. (2022) ‘Artificial intelligence and jobs: Evidence from online vacancies’, Journal of Labor Economics, 40(S1), pp. S293–S340. https://doi.org/10.1086/718327

Aghion, P., Jones, B.F. and Jones, C.I. (2017) ‘Artificial intelligence and economic growth’, NBER Working Paper 23928. Working paper. https://doi.org/10.3386/w23928

Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J. and Mané, D. (2016) ‘Concrete problems in AI safety’, arXiv preprint, arXiv:1606.06565v2. https://arxiv.org/abs/1606.06565v2

Bloom, N., Jones, C.I., Van Reenen, J. and Webb, M. (2020) ‘Are ideas getting harder to find?’, American Economic Review, 110(4), pp. 1104–1144. https://doi.org/10.1257/aer.20180338

Brynjolfsson, E., Hui, X. and Liu, M. (2019) ‘Does machine translation affect international trade? Evidence from a large digital platform’, Management Science, 65(12), pp. 5449–5460. https://doi.org/10.1287/mnsc.2019.3388

Brynjolfsson, E., Rock, D. and Syverson, C. (2021) ‘The productivity J-curve: How intangibles complement general purpose technologies’, American Economic Journal: Macroeconomics, 13(1), pp. 333–372. https://doi.org/10.1257/mac.20180386

Chen, M. et al. (2021) ‘Evaluating large language models trained on code’, arXiv preprint, arXiv:2107.03374v2. https://arxiv.org/abs/2107.03374v2

Fawzi, A. et al. (2022) ‘Discovering faster matrix multiplication algorithms with reinforcement learning’, Nature, 610, pp. 47–53. https://doi.org/10.1038/s41586-022-05172-4

Felten, E., Raj, M. and Seamans, R. (2021) ‘Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses’, Strategic Management Journal, 42(12), pp. 2195–2217. https://doi.org/10.1002/smj.3286

Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D. and Steinhardt, J. (2021) ‘Measuring massive multitask language understanding’, International Conference on Learning Representations. https://arxiv.org/abs/2009.03300v3

Hernandez, D. and Brown, T.B. (2020) ‘Measuring the algorithmic efficiency of neural networks’, arXiv:2005.04305. Preprint. https://arxiv.org/abs/2005.04305v1

Hoffmann, J. et al. (2022) ‘Training compute-optimal large language models’, arXiv preprint, arXiv:2203.15556v1. https://arxiv.org/abs/2203.15556v1

Jones, C.I. and Tonetti, C. (2020) ‘Nonrivalry and the economics of data’, American Economic Review, 110(9), pp. 2819–2858. https://doi.org/10.1257/aer.20191330

Jumper, J. et al. (2021) ‘Highly accurate protein structure prediction with AlphaFold’, Nature, 596, pp. 583–589. https://doi.org/10.1038/s41586-021-03819-2

Kaplan, J. et al. (2020) ‘Scaling laws for neural language models’, arXiv:2001.08361. Preprint. https://arxiv.org/abs/2001.08361v1

Li, Y. et al. (2022) ‘Competition-level code generation with AlphaCode’, Science, 378(6624), pp. 1092–1097. https://doi.org/10.1126/science.abq1158

Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I.D. and Gebru, T. (2019) ‘Model cards for model reporting’, Proceedings of the Conference on Fairness, Accountability, and Transparency, pp. 220–229. https://doi.org/10.1145/3287560.3287596

Ouyang, L. et al. (2022) ‘Training language models to follow instructions with human feedback’, arXiv preprint, arXiv:2203.02155v1. https://arxiv.org/abs/2203.02155v1

Parasuraman, R., Sheridan, T.B. and Wickens, C.D. (2000) ‘A model for types and levels of human interaction with automation’, IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30(3), pp. 286–297. https://doi.org/10.1109/3468.844354

Raji, I.D., Smart, A., White, R.N., Mitchell, M., Gebru, T., Hutchinson, B., Smith-Loud, J., Theron, D. and Barnes, P. (2020) ‘Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing’, Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pp. 33–44. https://doi.org/10.1145/3351095.3372873

Silver, D. et al. (2017) ‘Mastering the game of Go without human knowledge’, Nature, 550, pp. 354–359. https://doi.org/10.1038/nature24270