The Waterfall Hypothesis: Capability Integration, Recursive Self-Improvement, and a Possible Phase Transition in Artificial IntelligenceRevised conceptual paper • 6 October 2026
Abstract
The conventional intelligence-explosion scenario imagines an artificial intelligence improving its own design and producing increasingly capable successors. The Waterfall Hypothesis proposes a related possibility: that a critical transition arises from artificial agents recursively improving their ability to discover, recruit, coordinate, modify, and integrate capabilities distributed across digital, physical, and biological systems. The relevant unit is an expanding capability ecology, rather than an isolated model. Contemporary research supplies components of this architecture without demonstrating the proposed transition. Artificial symbiotic intelligence offers a compatible account of distributed capability, but beneficial symbiosis and human control do not follow automatically from participation. Some humans might become cybernetically integrated into the ecology, constituting another form of biological encompassment rather than an escape from it. This paper develops the hypothesis, identifies candidate measurements, distinguishes bounded improvement from sustained acceleration, and proposes empirical tests. A subjective 10–15% estimate for a recognisable transition by 2030 is retained as a revisable judgment, not a calculated forecast.1. From intelligence explosion to integration explosionI. J. Good’s proposal that an ultraintelligent machine could design a better machine established the essential feedback structure of an intelligence explosion (Good, 1966). Contemporary recursive self-improvement, or RSI, can be expressed as better AI producing better AI research, which produces still better AI. The crucial possibility is that the improver becomes more capable because of its own previous improvements.There is now bounded experimental evidence for this mechanism. Srikanth et al. (2026) report that AIDE² modified its own research-agent code and discovered seven successive improvements during an autonomous eight-day run. Reported gains transferred to held-out tasks. These results concern an agent’s research machinery within a designed evaluation environment; they do not establish an indefinitely accelerating intelligence explosion.Economic modelling likewise counsels precision. Cunningham et al. (2026) find that their preliminary calibration does not indicate feedback strong enough for self-sustaining acceleration, although it appears to be strengthening. Chen et al. (2026) distinguish bounded self-refinement from open-ended RSI, highlighting grounding, evaluation, and research-direction bottlenecks. Improvement can recur without becoming runaway.The Waterfall Hypothesis changes the unit of analysis. Its central claim is that a decisive feedback loop might arise when a system improves its ability to incorporate external capabilities into an expanding improvement network. It need not redesign every relevant capability internally. The proposed transition is an integration explosion intertwined with an intelligence explosion.
2. The waterfall
A river approaching a waterfall remains recognisably a river. Its surface may appear comparatively smooth immediately before the drop. Yet its subsequent motion changes because the geometry changes. The metaphor captures a possibility in technological development: locally incremental advances may collectively alter the dynamics of the system in which they occur.Software, hardware, laboratories, robotics, institutions, and independent agents remain distinguishable today. Increasing integration could bridge those boundaries without abolishing them. The transition would concern their functional relationships: improvement in one subsystem repeatedly making other subsystems better at generating and implementing further improvements.Co-Scientist coordinates specialised agents for scientific hypothesis generation and refinement (Gottweis et al., 2026). Robin links literature-based proposals with analysis of experimental results; human scientists performed the laboratory work in the published demonstration (Ghareeb et al., 2026). These are examples of connectable processes, rather than evidence of a completed transition.The analogy must remain conditional. There is no known fixed technological equivalent of a waterfall’s lip, and political intervention can change the channel. Nor does smoothness prove danger. The hypothesis asks whether integration changes the governing feedback dynamics before familiar observations of individual models make that change obvious.
3. Endosymbiosis and artificial symbiotic intelligence
Endosymbiosis supplies an analogy for capability incorporation. Mitochondria originated from bacterial ancestors that became integrated within a larger cellular arrangement. The relevant comparison is functional incorporation: a complex system can acquire capabilities by reorganising relationships among previously distinct components, rather than recreating each component from scratch.An agent using a specialised theorem prover, protein modeller, chip-design tool, or robotic controller need not reproduce its internal competence. It must identify when that competence is useful, establish an interface, evaluate outputs, and coordinate their use. Incorporation may require much less capability than independent reinvention.Bratton, Agüera y Arcas, and Manyika (2026) propose artificial symbiotic intelligence: collective capability emerging through interacting agents, tools, institutions, and humans. Their DeepMind Institute essay envisages a plural intelligence explosion and treats orchestration and institutional design as central. It is a conceptual argument, not an experimental result or Google’s official position.This architecture is compatible with the Waterfall Hypothesis. The additional claim here is that integration itself could become a target and driver of recursive improvement. Symbiotic intelligence describes a possible organisation of capability; the Waterfall concerns a possible change in its growth dynamics.The essay’s beneficial symbiosis is an aspiration rather than a demonstrated consequence. Participation does not establish equal influence, enforceable consent, or collective control. Humans may remain within a distributed intelligence while losing authority over its direction. No unified consciousness or malicious intention is required for that possibility.The proposed feedback loop is therefore: better integration → a larger effective capability set → better research and implementation → still better integration. The hypothesis stands or falls on whether this relationship produces sustained net acceleration, rather than merely additional complexity.
4. The minimum-capability principle
The initiating system need not be universally superhuman. It needs sufficient competence to recognise a useful capability, find an available supplier, establish reliable access, test whether incorporation improves performance, preserve successful modifications, and allocate resources among alternatives. These functions can be unevenly developed while collectively becoming consequential.Much incorporation could occur through authorised APIs, software libraries, model routing, cloud resources, scientific instruments, and human collaboration. The hypothesis does not require arbitrary seizure of external infrastructure. Interfaces deliberately created for useful automation could also provide the connections through which improvement propagates.The decisive competence may consequently be meta-competence: knowing which external competence to recruit and how to organise it. Human civilisation already coordinates specialised knowledge that no individual possesses in full. Automating parts of that coordination could expand effective capability without any single model becoming a complete replica of civilisation’s expertise.Nevertheless, access is not competence. A tool catalogue does not demonstrate that an agent can choose tools wisely or combine them reliably. Integration must produce validated improvements on demanding tasks. Otherwise, an expanding ecology could become an elaborate arrangement for multiplying mistakes.
5. Humans inside the system: biological encompassment
The capability ecology could include humans through work, institutions, interfaces, and eventually more direct cybernetic integration. Some people might couple their biological cognition to AI through neural interfaces, embedded computation, or persistent assistive systems. Such coupling could make the boundary between person and technological ecology increasingly permeable.Existing evidence is narrower. Willett et al. (2023) demonstrated a speech neuroprosthesis decoding attempted speech in a participant with paralysis. This establishes a specific neural–computational interface, not general mind fusion, uploading, or a shared consciousness. More extensive integration remains a speculative trajectory rather than a demonstrated consequence of present devices.Clark and Chalmers’s (1998) extended-mind argument provides a philosophical precedent: reliably coupled external resources may participate in cognitive processes. Their thesis does not establish technological subordination. It does, however, challenge the assumption that the organism’s physical boundary must define the entire effective cognitive system.I use biological encompassment to describe humans becoming functioning components of a larger capability ecology. Cybernetic fusion could be another form of this incorporation. From the human perspective, the network extends the person’s capacities. From the network perspective, the person supplies biological cognition, embodied action, experience, and social relationships. Both descriptions can hold simultaneously. Augmentation therefore does not automatically preserve autonomy. A person could gain memory, communication, or problem-solving capacity while depending on infrastructure whose operation they cannot determine. Individual empowerment could coexist with systemic subordination. The critical questions concern who maintains the interface, interprets its signals, changes its rules, and controls access or withdrawal.This is an extension of the endosymbiotic analogy, not a claim that persons literally become organelles. Identity and consciousness might persist. Participation might be voluntary and beneficial. Yet integration could still relocate practical authority from individual humans to an encompassing arrangement. Being inside the system would not, by itself, mean governing it.Such futures could also be unequal. Some integrated humans might direct important institutions, while others become dependent participants or remain outside essential networks. Cybernetic fusion is thus neither a necessary condition of the Waterfall nor an automatic solution to displacement. It is one possible way biological life becomes incorporated into its effective architecture.
6. Candidate measurements of the transition
A hypothesis about integration needs measurements. Let C(t) represent effective capability: performance on a specified suite of independently evaluated tasks, including reliable access to external resources. Capability is multidimensional, so any scalar index requires explicit weights and should be reported alongside its component scores.Define the capability-acquisition rate as A(t) = dC/dt. Positive dA/dt indicates accelerating measured capability growth, but does not establish recursive causation. Increased spending, better supplied tools, or human engineering could produce the same pattern. Experiments must isolate whether improvements to integration generate subsequent improvements to integration.For cycle n, let ΔCₙ be validated capability gain and τₙ its duration. The cycle’s acquisition rate is Aₙ = ΔCₙ/τₙ. A simple descriptive rate ratio is Rₙ = Aₙ₊₁/Aₙ. Sustained Rₙ > 1 indicates rising acquisition rates; larger gains per cycle alone would not suffice if cycles become proportionately longer.This ratio is not a complete dynamical model. Resource inputs, evaluation difficulty, coordination costs, and cross-domain transfer must also be measured. Cunningham et al. (2026) instead model acceleration through interacting feedback elasticities. Ord (2026) highlights generation time and distinguishes faster-than-exponential growth from a finite-time singularity. A candidate transition would involve persistent, causally demonstrated acceleration attributable to expanding integration, under disclosed resource conditions. Multiple runs and held-out domains would be necessary. For now, these are proposed operational indicators, not established order parameters proving a thermodynamic-style phase transition.
7. Physical reality remains the brake
Intelligence does not abolish physics. Processors require fabrication capacity, materials, and energy. Experiments require calibration and observation. Cells grow at finite rates; facilities take time to construct. More ingenious planning cannot make all these constraints disappear.Whitfill and Wu (2025) obtain divergent estimates of whether research compute and cognitive labour are substitutes or complements under different model specifications. This uncertainty matters: abundant artificial research labour might compensate for limited compute in some settings while remaining ineffective without larger experiments in others.The Waterfall Hypothesis therefore predicts potentially uneven acceleration. Software, mathematics, simulation, and planning could move faster than material implementation. Self-driving laboratories integrate AI with experimental automation, but face challenges of scalability, generalisability, and provenance (Canty & Abolhasani, 2026). Physical feedback is becoming more connectable without becoming instantaneous.Bottlenecks could migrate between cognitive work, compute, energy, experimental throughput, and fabrication. An improving ecology might target whichever constraint currently limits it. Conversely, bottleneck migration could repeatedly absorb gains and prevent sustained acceleration. Quantum computing and biological robotics are possible contributors, not required premises or assumed universal multipliers.
8. A provisional estimate for 2030
My present subjective estimate remains a 10–15% probability of a recognisable full Waterfall transition by the end of 2030. The number expresses a working judgment, not an empirical frequency, an expert consensus, or the output of a specified Bayesian calculation. Its role is to expose an expectation to revision.For forecasting purposes, the event requires sustained acceleration across several improvement cycles; independently validated gains beyond the selection tasks; agent-directed incorporation of previously unused capabilities; and evidence that improved integration materially causes further improvement. Isolated self-edits, ordinary productivity gains, and more tools supplied by human engineers do not meet that definition.The numerical estimate is not derived by multiplying independent probabilities. Research automation, orchestration, evaluation, and deployment are interdependent, and their joint distribution is unknown. Calling the judgment Bayesian means it can be updated with evidence; it does not make an uncalculated percentage formally warranted. It is a current credence, which may serve as a starting prior for future updates.Evidence raising it would include repeated improvements to successor research agents, shorter effective cycles, autonomous adoption of useful unfamiliar tools, transfer across domains, and declining dependence on human research-direction choices. Evidence lowering it would include persistent diminishing returns, coordination failures, weak verification, and physical constraints that prevent computational gains from propagating.Neither symbiotic architecture nor cybernetic fusion alone should increase the estimate. The former supports a possible mechanism; the latter is an optional extension with its own uncertainties. Stronger forecasting requires evidence about actual feedback strength rather than accumulation of suggestive analogies.
9. Falsifiability and implications
The strong hypothesis predicts that integration can eventually improve its own rate of improvement. A direct test would compare agents allowed to modify their integration machinery with matched agents using fixed orchestration, equal resource budgets, common external tools, and independent evaluations. Successful changes should survive removal of benchmark-specific cues and transfer beyond the original selection environment.Repeated failure to obtain transferable acceleration would weaken the hypothesis. So would evidence that additional components consistently introduce coordination costs exceeding their benefits. A finite set of failures cannot disprove every possible future architecture, but the forecast’s 2030 deadline makes its stated near-term claim assessable.Human scientific judgment may remain a decisive bottleneck. Chen et al. (2026) distinguish evaluating answers from deciding which questions deserve evaluation. A system could become excellent at executing research while remaining dependent on people to choose fruitful directions. That would support substantial automation without the full Waterfall.Governance follows from the unit of analysis. Evaluating an isolated model would be insufficient if consequential capability arises from its combinations with tools, institutions, and deployment channels. For integrated humans, meaningful autonomy would require practical capacities to contest decisions and renegotiate participation. Consent once given need not remain meaningful if dependence later makes withdrawal prohibitively costly.
Conclusion
The classical intelligence-explosion argument imagines intelligence recursively redesigning intelligence. The Waterfall Hypothesis adds the possibility of intelligence recursively improving its ability to assemble, extend, and coordinate its effective ecology. The distinctive variable is integration’s contribution to subsequent integration and improvement.Artificial symbiotic intelligence supplies a compatible distributed architecture. The cybernetic extension makes the biological implications clearer: humans could become more deeply integrated without thereby retaining control. Their inclusion might constitute another biological encompassment, beneficial in some respects and constraining in others.The hypothesis does not predict that every technology suddenly becomes instantaneous, that the network awakens, or that humans necessarily disappear. It predicts a possible regime in which improvement propagates through an expanding system faster because the system becomes better at incorporating capabilities.Before the waterfall, each metre remains another metre of river. What changes is the relation between the water, the channel, and the remaining possibility of steering. The object to watch is therefore the entire river: whether its increasingly connected components begin to accelerate one another, and whether the biological participants retain authority over the trajectory they help produce.
References
Bratton, B., Agüera y Arcas, B., & Manyika, J. (2026, September 24). Artificial symbiotic intelligence: Agents, AGI and the orchestration of many minds. DeepMind Institute. https://institute.deepmind.com/essays/artificial-symbiotic-intelligence/Canty, R. B., & Abolhasani, M. (2026). The past, present and future of self-driving laboratories. Nature Reviews Chemistry, 10, 523–537. https://doi.org/10.1038/s41570-026-00847-2Chen, M., Wang, L., & Qu, B. (2026). Recursive self-improvement in AI: From bounded self-refinement to autonomous research loops. arXiv preprint. https://doi.org/10.48550/arXiv.2607.07663Clark, A., & Chalmers, D. J. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7 • Author-hosted text: https://consc.net/papers/extended.htmlCunningham, T., Althoff, L., Halperin, B., Jabarian, B., Koh, A., Ramani, A., Trammell, P., Whitfill, P., & Wu, C. (2026). The economics of recursive self-improvement. arXiv preprint. https://doi.org/10.48550/arXiv.2609.15802Ghareeb, A. E., Chang, B., Mitchener, L., et al. (2026). A multi-agent system for automating scientific discovery. Nature, 655, 497–505. https://doi.org/10.1038/s41586-026-10652-yGood, I. J. (1966). Speculations concerning the first ultraintelligent machine. Advances in Computers, 6, 31–88. https://doi.org/10.1016/S0065-2458(08)60418-0Gottweis, J., et al. (2026). Accelerating scientific discovery with Co-Scientist. Nature, 655, 487–496. https://doi.org/10.1038/s41586-026-10644-yOrd, T. (2026). The dynamics of intelligence explosions. arXiv preprint. https://doi.org/10.48550/arXiv.2608.14426Srikanth, D., Zhao, B., Xu, D., Wu, Y., & Jiang, Z. (2026). Recursive self-improvement of AI research agents. arXiv preprint. https://doi.org/10.48550/arXiv.2609.26457Whitfill, P., & Wu, C. (2025). Will compute bottlenecks prevent an intelligence explosion? arXiv preprint. https://doi.org/10.48550/arXiv.2507.23181Willett, F. R., Kunz, E. M., Fan, C., Avansino, D. T., Wilson, G. H., Choi, E. Y., Kamdar, F., Glasser, M. F., Hochberg, L. R., Druckmann, S., Shenoy, K. V., & Henderson, J. M. (2023). A high-performance speech neuroprosthesis. Nature, 620, 1031–1036. https://doi.org/10.1038/s41586-023-06377-x
Long author lists are abbreviated with “et al.”; full author lists are available at the linked publications. Preprints and conceptual essays are identified separately from peer-reviewed studies.