Potential Pathways to Superintelligence

Automated AI research (AI R&D automation) is the use of AI systems to perform some or all of the activities of AI research and development: generating ideas, writing and running experiments, analysing results, engineering software and writing up findings. It is of interest because, if AI systems could improve AI systems, progress might accelerate; whether and how far that could happen is Hypothetical or Contested, and the evidence so far concerns partial automation in constrained settings.

The intelligence-explosion argument (ASI-EXP-001) rests on a feedback loop in which AI contributes to the research that produces better AI. Automated AI research is the empirical counterpart of that argument: it asks how much of the research process current systems can do, how that compares with human researchers, and what limits remain. The topic is also relevant to safety policy; the RE-Bench authors state that AI R&D automation is a key capability for safety policies to anticipate, while noting that few realistic benchmarks compare AI systems with human experts (Wijk et al., 2024).

This article distinguishes four things that are often run together.

1. Automating engineering tasks inside AI R&D, such as writing training code or optimising kernels.
2. Automating the research loop, from hypothesis to paper, in a limited domain.
3. Automating algorithm discovery, where a system searches for new algorithms against an automatic evaluator.
4. Closing the loop, in which the products of automated research improve the system that does the research. No source reviewed shows this at scale.

Early proposals for self-improving machines (Good, 1966) were theoretical. Automated machine-learning methods, which search over architectures or hyperparameters, preceded the current wave; they automate narrow parts of the process and are not the focus here. Hutter (2012) analysed under what conditions an intelligence explosion would occur, showing that conclusions depend on assumptions about returns to intelligence.

From 2023 onwards, benchmarks and systems appeared that target research tasks directly. MLAgentBench (Huang et al., 2023) evaluates language-model agents on machine-learning experimentation tasks. MLE-bench (Chan et al., 2024) uses Kaggle competitions. RE-Bench (Wijk et al., 2024) compares agents with human experts on research-engineering environments. System-level work includes The AI Scientist (Lu et al., 2024), which the authors describe as a framework in which language models carry out the full research cycle, with a stated cost of under USD 15 per paper in their experiments, and its successor (Yamada et al., 2025). Algorithm-discovery systems include AlphaEvolve (Novikov et al., 2025), and AI co-scientist (Gottweis et al., 2025) targets hypothesis generation in biomedicine.

#4.1 A model of the feedback loop

A simple stylised model illustrates why assumptions matter. Let $A(t)$ be the level of AI capability and suppose the rate of algorithmic and software progress depends on the research effort supplied by AI systems:

where $R(A)$ is the effective research labour supplied (increasing in $A$) and $C$ is the other inputs, such as compute and experimental capacity, that research needs. If $C$ is not limiting and returns to research labour are high enough, progress accelerates; if $C$ is limiting, or if ideas become harder to find, the feedback is damped. Bloom et al. (2020) document that research productivity has declined in several fields in the data they study, which is an argument that returns to research effort diminish. Erdil and Besiroglu (2023) review arguments for explosive growth from AI automation and weigh counterarguments including bottlenecks. Ho et al. (2024) estimate, for language models on two benchmarks from 2012 to 2023, that the compute needed to reach a fixed performance halved approximately every 8 months (95% confidence interval roughly 5 to 14 months), and conclude that increased compute contributed more to performance gains than algorithmic innovations did. Together, this suggests that compute is a major input, which matters for how far software-only research automation could go; the question is not settled by these sources.

This model is a schematic, not a measured system; $f$ and $R$ are not known.

#4.2 What the benchmarks measure

RE-Bench. Seven open-ended research-engineering environments with data from 71 8-hour attempts by 61 distinct human experts; 82% of expert attempts achieved a non-zero score and 24% matched or exceeded the strong reference solutions. The best AI agents achieve a score 4x higher than human experts when both are given a total time budget of 2 hours per environment, whereas humans achieve 2x the score of the top AI agent when both are given 32 total hours. The authors also report that agents can generate and test solutions over ten times faster than humans at much lower cost (Wijk et al., 2024). The result describes seven environments and the models tested at that date.

MLE-bench. 75 ML-engineering competitions from Kaggle with human baselines from public leaderboards. The best-performing setup at the time, OpenAI's o1-preview with AIDE scaffolding, achieved at least the level of a Kaggle bronze medal in 16.9% of competitions (Chan et al., 2024).

Time horizons. Kwa et al. (2025) measure the length of software tasks completed with 50% success. They report a 50% time horizon of around 50 minutes for models such as Claude 3.7 Sonnet and a doubling time of approximately seven months since 2019. The paper includes RE-Bench tasks among its task sources. The metric covers software tasks and a specified success rate.

Interpretation. These instruments show that current agents can do parts of research engineering quickly and cheaply and that the advantage is greatest with short time budgets. They do not show that agents can conduct the open-ended, long-horizon work of choosing research directions, which none of the benchmarks measures directly.

#4.3 End-to-end research systems

The AI Scientist (Lu et al., 2024) generates ideas, writes and runs code, visualises results, drafts a paper and reviews it with a simulated reviewer; the authors tested it in diffusion modelling, transformer language modelling and learning dynamics, and report that their automated reviewer performs close to human reviewers and that some generated papers pass the acceptance threshold of a top machine-learning conference by the automated reviewer's standard. The reviewer is itself an automated system built by the authors, so this is not independent evidence of publishable quality. The AI Scientist-v2 (Yamada et al., 2025) uses agentic tree search and a vision-language feedback loop; the authors submitted three fully autonomous papers to an ICLR workshop, one of which scored above the average human acceptance threshold. A workshop acceptance threshold is lower than that of a main conference, and the result concerns a single submission batch. Both are Preliminary: results come from the developing group and a small number of cases.

#4.4 Algorithm discovery

AlphaEvolve (Novikov et al., 2025) is an evolutionary coding agent in which language models rewrite code guided by automatic evaluators. The authors apply it to computing infrastructure, open scientific problems and algorithm design, and report a procedure for multiplying two 4x4 complex-valued matrices using 48 scalar multiplications, which they describe as the first improvement after 56 years over Strassen's algorithm in this setting. These systems work where an automatic evaluator can score candidate solutions, which is a narrower condition than most research satisfies. The AI co-scientist (Gottweis et al., 2025) is an example of a multi-agent system for hypothesis generation; its hypotheses are tested separately from the generation system. [VERIFICATION NEEDED: how the paper validates hypotheses was not checked beyond the title and listing] See APP-SCI-001.

The components of a research-automation system are typically a language-model core, an agent scaffold (planning, tool use, code execution), an evaluation harness and a memory or archive of prior attempts. Their general technical background is in TECH-AGENT-001.

• Research engineering assistance. Agents complete portions of software and ML-engineering tasks, as in RE-Bench and MLE-bench, which are evaluation settings and not measures of deployed productivity.
• Algorithm and kernel discovery. AlphaEvolve is described by its authors as applied to Google's computing infrastructure and to mathematical problems; the matrix-multiplication result is one reported example (Novikov et al., 2025).
• Paper-generation pipelines. The AI Scientist and its successor are laboratory systems (Lu et al., 2024; Yamada et al., 2025).
• Scientific hypothesis generation. The AI co-scientist is a laboratory system for biomedical hypotheses (Gottweis et al., 2025).

Deployment inside AI developers is not documented in the sources reviewed, and the extent of internal use of AI in AI development at specific companies is, for the purposes of this article, undisclosed.

Potential advantages. Speed and cost (agents generate and test solutions more than ten times faster than humans at much lower cost in RE-Bench, per the authors), parallelism and availability.

Limitations.

• Short-horizon strength, long-horizon weakness. Agents lead at 2-hour budgets and trail at 32 hours in RE-Bench (Wijk et al., 2024).
• Benchmarks differ from frontier research. Environments are well specified and scored; frontier research involves choosing problems and judging significance.
• Evaluator dependence. Systems such as AlphaEvolve need automatic evaluators; results in domains without them are not documented in the sources reviewed.
• Self-grading. End-to-end systems partly rely on automated reviewers built by the same group.
• Compute and experiment bottlenecks. Progress depends on compute for experiments (Ho et al., 2024; Erdil & Besiroglu, 2023). A system that proposes ideas faster than they can be tested faces a limit.
• Diminishing returns. Bloom et al. (2020) argue ideas are getting harder to find in the data they study; whether this applies to AI research is uncertain.
• Safety implications. Rapid capability growth would shorten the time available for evaluation and oversight; see SAFE-EVAL-001 and SAFE-OVERSIGHT-001.

Open questions.

1. How much of the AI research pipeline is bottlenecked by compute rather than by ideas or labour?
2. Can agents do research that requires judgement over weeks to months?
3. Is there a threshold at which the feedback becomes self-sustaining, or does it saturate? Contested.
4. How should automation of AI R&D be measured in ways that are hard to game?

As of 2026-10-15, the sources reviewed show partial automation of research engineering and demonstrations of end-to-end research systems in narrow settings. They do not show closed-loop recursive self-improvement, and the claim that recursive self-improvement leads to an intelligence explosion is Hypothetical and Contested (see ASI-EXP-001). Benchmark values quoted above date from 2024 and 2025 papers and apply to the models tested then; later systems may differ, and current values should be taken from the Living Reference (MODEL-001, FACT-001) rather than from this article. Evaluations from developers and independent groups use different protocols, so comparisons across them require care.

Observed trends (not forecasts): the benchmark results above show increasing capability on research-engineering tasks between the cited papers' model generations, and the time-horizon metric reports a doubling time of about seven months in its data (Kwa et al., 2025). Extrapolating a trend into the future is a forecast and is registered below.

Near term (0 to 3 years). Further benchmarks of AI R&D capability and agent evaluations; continued system-level demonstrations. Medium term (3 to 10 years). The central uncertainty is whether automation of research engineering extends to research direction-setting and whether compute limits it. The sources reviewed do not resolve this. Long term (10+ years). Scenarios in which AI research is largely automated are Hypothetical; see ASI-PATH-001.

#Registered forecast

• Foundational Concepts:
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: The hypothesis that automated research motivates.
  • TECH-AGENT-001 Autonomous AI Agents: Agent architectures used in research automation.
  • ASI-001 Artificial Superintelligence: Overview of ASI.
• Related Methods & Technologies:
  • AGI-TEST-001 Testing Claims of General Intelligence: Validity issues for agentic benchmarks.
  • CAP-AGENTEVAL-001 Evaluating Autonomous Agents: Agent evaluation methods.
  • APP-SCI-001 AI in Science: Scientific applications.
  • ECON-COMP-001 The Economics of Compute & Scaling: Compute as a limiting input.
• Advanced Implications & Future Trajectories:
  • ASI-PATH-001 Pathways to Superintelligence: Where automated research sits among pathways.
  • SAFE-EVAL-001 Dangerous Capability and Alignment Evaluations: Evaluating AI R&D capability as a risk indicator.
  • SAFE-OVERSIGHT-001 Scalable Oversight and Corrigibility: Oversight under faster capability growth.

Pathways to superintelligence are the proposed technical and social routes by which an intellect greatly exceeding human performance might come to exist. The best-known classification, from Bostrom (2014), distinguishes artificial intelligence, whole brain emulation, biological cognitive enhancement, brain-computer interfaces, and networks and organisations. All of these remain Hypothetical: no route has produced a system that meets a general definition of superintelligence (see ASI-DEF-001).

AGI-PATH-001 covers routes to human-level machine intelligence, and ASI-EXP-001 covers the hypothesis that capability could grow rapidly once certain thresholds are passed. This article is concerned with the intermediate question: given human-level systems, or without them, by what mechanisms might capability go substantially beyond the human range, and what would each mechanism require?

The question matters because the routes differ in the following respects, which bear on safety and governance.

• Predictability. A route that extends known scaling trends is easier to anticipate than one that requires a conceptual breakthrough.
• Speed and discontinuity. Routes differ in whether progress is likely to be smooth, stepwise or abrupt.
• Controllability. Whether the resulting system's goals and behaviour can be inspected and shaped differs between routes. For example, an emulation of a human brain would inherit human-like psychology in some respects, while an artificial system trained by optimisation may not.
• Actors and infrastructure. Different routes require different resources, which determines who could pursue them (see ECON-INFRA-001).

Beyond classification, this article identifies what evidence bears on each route and which claims are matters of argument rather than data.

Good (1966) raised the possibility that a machine surpassing human intellect could improve its own design. Bostrom (1998) discussed estimates of the brain's processing power, when hardware might match it, and how quickly superintelligence could follow human-level AI; the paper's abstract argues that superhuman AI is likely within the first third of the twenty-first century [VERIFICATION NEEDED: wording of the timing claim taken from a secondary summary of the abstract]. No system meeting the definition exists as of this review, and the date is mentioned only as an example of a timing claim in this literature.

Sandberg and Bostrom (2008) produced a roadmap for whole brain emulation, the replication of a brain's function in a computer, following a workshop at the Future of Humanity Institute in Oxford in 2007. Bostrom (2014) then set out the classification used here and analysed the dynamics of a transition. In the machine-learning era, the empirical discussion turned to scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022), to self-play systems that exceeded human performance without human game data (Silver et al., 2017), and to the question of whether AI systems could take part in AI research itself (see ASI-AIRD-001).

#4.1 Pathway A: Artificial intelligence

This is the route most discussed in current research and policy. It has several mechanisms, which can combine.

A1. Scaling of learning systems. Kaplan et al. (2020) reported that the cross-entropy loss of language models falls as a power law with model size, dataset size and training compute, over the ranges they studied. Hoffmann et al. (2022) reported, for compute-optimal training, that model size and training tokens should be scaled in roughly equal proportion. These are empirical regularities for a loss metric on a given model family, and extrapolating them to superhuman capability is an inference, not a result. A stylised form is

with $N$ parameters, $D$ tokens and fitted constants $E, A, B, \alpha, \beta$; the constants and functional form are from the fitted literature and differ across studies. Whether lower loss translates into capabilities beyond human experts across domains is Contested.

A2. Algorithmic progress. Ho et al. (2024) analysed more than 200 language-model evaluations on two benchmarks from 2012 to 2023 and estimated that the compute needed to reach a fixed performance threshold halved approximately every 8 months, with a 95% confidence interval of roughly 5 to 14 months. They also concluded that increased compute contributed more to overall performance gains than algorithmic innovations did, and noted that noisy benchmark data limits the precision of the estimates. These figures concern those benchmarks and that period; they are not a forecast of future algorithmic efficiency.

A3. Self-play and search. Silver et al. (2017) reported a Go system trained by reinforcement learning from self-play without human game data that outperformed earlier human-data-based systems. This demonstrates a mechanism by which performance beyond the best humans can be reached in a domain with a fixed, perfectly specified objective and cheap simulation. Whether the mechanism transfers to open-ended domains without such an objective is Contested and depends on the availability of reliable verification.

A4. Automated AI research. If AI systems become able to do research that improves AI systems, feedback could accelerate progress. This is the mechanism behind the intelligence-explosion hypothesis; its evidence status is discussed in ASI-AIRD-001 and ASI-EXP-001. Erdil and Besiroglu (2023) review arguments for explosive economic growth from AI automation, identify drivers (scalable AI systems, a fast increase in the number of AI workers, and a sharp jump in output from automation) and assess counterarguments such as regulation and production bottlenecks. They conclude that explosive growth is plausible if AI can largely replace human labour, but that high confidence is not yet justified. Hutter (2012) analysed conditions under which an intelligence explosion would or would not occur, in a mathematical framework; the conclusions depend on assumptions about how intelligence translates into the ability to improve itself.

A5. Amplified supervision and oversight. Christiano et al. (2018) proposed iterated amplification, a training scheme for tasks that humans cannot directly evaluate, in which a weak expert is amplified by help from copies of the learning system. Burns et al. (2023) test weak-to-strong generalisation as an analogy for supervising systems beyond human ability. These are not routes to superintelligence in themselves but research on whether a system trained under such schemes can safely be capable beyond its supervisors (see SAFE-OVERSIGHT-001).

#4.2 Pathway B: Whole brain emulation

Sandberg and Bostrom (2008) describe emulation as obtaining a detailed scan of a brain and implementing its function in a computer, and discuss the required technologies: scanning, modelling the neural substrate and computation. The report was written to examine feasibility rather than to predict a date. Emulation would yield human-like minds, possibly run faster or copied; this corresponds to the speed or collective forms in Bostrom (2014). The approach is bounded by the state of neuroscience and scanning technology, and no human brain has been emulated. As of the sources reviewed, this pathway is Hypothetical, and its relation to AI progress is debated: some argue that AI will arrive first. The sources reviewed contain no evidence settling that question.

#4.3 Pathway C: Biological cognitive enhancement

Genetic selection, genetic modification, pharmacology and training could in principle raise human cognitive ability. Bostrom (2014) discusses this route and argues that biological enhancement is likely to be slow relative to machine routes because of generation times, though the discussion is argumentative and the claim is Contested. [VERIFICATION NEEDED: the specific chapter and wording of Bostrom's comparison of biological and machine routes were not re-checked]. Ethical and legal constraints are significant and outside the scope of this article.

#4.4 Pathway D: Brain-computer interfaces

Interfaces between brains and computers could extend human cognition. The sources reviewed for this article do not provide evidence that current interfaces enhance general cognition beyond the human range; this pathway is Hypothetical.

#4.5 Pathway E: Networks and organisations

Collective intelligence arises from organisations, markets and the internet. Bostrom (2014) includes the strengthening of such networks as a route to collective superintelligence. Human institutions already outperform individuals at many tasks, which raises the question, treated in ASI-DEF-001, of how the reference class is defined.

Pathways are not exclusive. For instance, scaling, algorithmic progress and automated research may operate together, and Pathway A is generally treated as the principal route in current discussion because it is the one on which active research is concentrated.

There are no applications of superintelligence; the following are partial demonstrations of mechanisms.

• Superhuman in a narrow domain by self-play. Silver et al. (2017). A laboratory result in a fixed-rule game.
• Measured scaling behaviour. Kaplan et al. (2020); Hoffmann et al. (2022). Empirical regularities for language-model loss.
• Measured algorithmic efficiency. Ho et al. (2024), for language models on the benchmarks they studied.
• Oversight analogues. Burns et al. (2023) report experiments with weak supervisors and strong models as an analogue; Christiano et al. (2018) is a proposal for amplified supervision.
• Emulation. Sandberg and Bostrom (2008) is a roadmap and not an implementation.

Readers should not infer from any of these that a particular pathway is close to completion.

Value of the pathway framework. It structures a speculative field, identifies the bottlenecks specific to each route and clarifies which empirical observations bear on which claims.

Limitations.

• Pathways were classified before the current generation of AI systems, and the boundaries between them have shifted.
• Evidence for the machine-learning pathway concerns loss, benchmark performance and narrow domains, not general superhuman capability.
• Hardware, energy and data constraints may bind before algorithmic limits do. These are covered in ECON-INFRA-001 and ECON-COMP-001.
• The relevant physical and theoretical limits to intelligence are not known. Skeptical arguments, for example about diminishing returns to self-improvement, are discussed in ASI-EXP-001.

Open questions.

1. Does scaling plus algorithmic progress suffice to reach general superhuman capability, or are new ideas required? Contested.
2. Do verification-limited domains (research, strategy, social reasoning) permit the kind of self-play improvement seen in games?
3. Would automated AI research accelerate progress, or be limited by compute and experiments? See ASI-AIRD-001.
4. How would emulation or enhancement pathways interact with AI progress?

As of 2026-10-15, no pathway has produced a system meeting a general definition of superintelligence, and the sources reviewed for this article do not claim otherwise. The best-documented mechanisms are in the machine-learning pathway: loss scaling regularities (Kaplan et al., 2020; Hoffmann et al., 2022), measured algorithmic efficiency gains (Ho et al., 2024), and superhuman narrow performance via self-play (Silver et al., 2017). Whether these extend to general superhuman capability is Contested. Whole brain emulation, biological enhancement and brain-computer interfaces are Hypothetical, in the sense that no demonstration relevant to superintelligence is documented in the sources reviewed. For AGI timeline surveys see AGI-TIME-001; for quantitative current inputs (compute, model sizes, costs), consult the Living Reference (MODEL-001, FACT-001) because such figures change quickly.

This section is descriptive. It registers no forecast; dated forecasts, if any, belong in the Forecast Register (FCST-REG-001).

• Near term (0 to 3 years). Continued research on scaling, algorithmic efficiency, agentic systems and automated research, together with oversight research on supervising systems whose outputs are difficult to verify.
• Medium term (3 to 10 years). Evidence on whether verification-limited domains yield to the methods that succeeded in games and on whether automated research accelerates AI progress would discriminate between the AI mechanisms. The sources reviewed do not indicate how these questions will resolve.
• Long term (10+ years). Whole brain emulation and enhancement pathways cannot be assessed from the sources reviewed beyond their roadmaps. Scenarios about the aftermath of superintelligence are covered in PHIL-CIV-001.

• Foundational Concepts:
  • ASI-001 Artificial Superintelligence: Overview of ASI.
  • ASI-DEF-001 Defining and Measuring Superintelligence: What counts as superintelligence.
  • AGI-PATH-001 Pathways to AGI: The preceding stage of the argument.
• Related Methods & Technologies:
  • TECH-SCALE-001 Scaling Laws: Scaling regularities in detail.
  • ASI-AIRD-001 Automated AI Research: Mechanism A4.
  • ECON-INFRA-001 Energy, Data Centers and Semiconductor Supply Chains: Physical constraints.
  • SAFE-OVERSIGHT-001 Scalable Oversight and Corrigibility: Supervising systems beyond human ability.
• Advanced Implications & Future Trajectories:
  • ASI-EXP-001 The Intelligence Explosion Hypothesis: Takeoff dynamics.
  • SAFE-XRISK-001 AI Existential Risk & Catastrophic Scenarios: Risk implications.
  • PHIL-CIV-001 Civilization and Advanced Intelligence: Long-run scenarios.