Conscious incompetence, the conformity of large language models, and the coming geopolitics of epistemic capability

The strategic entwinement of innovation and state power has, since the Industrial Revolution, been the central grammar through which the international system has reorganised itself in successive epochs of technological transformation. From the rifled artillery of the mid-nineteenth century to the wireless telegraphy of the early twentieth, from the institutional fusion of science and state articulated in Vannevar Bush’s Science, the Endless Frontier (1945) to the moonshot architecture of the Defense Advanced Research Projects Agency, the relationship between technological emergence and geopolitical influence has not been incidental but constitutive. It is within this long durée — a durée whose contours have occupied the better part of my doctoral research on aerospace innovation ecosystems from the Cold War to the Global War on Terror — that the present essay situates a contention that is, I think, both timely and structurally underexamined: that the dominant paradigm of contemporary artificial intelligence is, by design, an instrument of consensus rather than discovery, and that the geopolitical consequences of this design choice are likely to prove decisive within the coming two decades.

The proposition advanced here is that we have, at staggering collective expense, constructed the most sophisticated mirror in history. The mirror reflects the centre of recorded human thought with extraordinary fluency. It does not, and by its present architecture cannot, illuminate the regions at the edges of that thought — the conscious-incompetence frontier where the next generation of strategic, scientific, and doctrinal breakthroughs sits latent in open literature, awaiting an ecosystem capable of recognising what it has. This essay makes the case for a different class of system, one I term Inverse Intelligence: an artificial intelligence architected not to optimise within the consensus distribution of a knowledge domain, but to map and probe its boundaries.

The case proceeds across five movements. First, a reconsideration of the biological metaphor that lends present systems their name. Second, a structural critique of the inductive paradigm. Third, a recovery of the conscious-incompetence framework as a tractable analytic frontier. Fourth, the historical case of the Ufimtsev–Lockheed asymmetry that retrospectively validates the proposed framework. And fifth, the integration of Inverse Intelligence into the dual-formula analytical architecture I have developed across the doctoral programme, with implications for the strategic horizon of 2050.

I. The Substrate Misnomer: Neurons, Glia, and the Caricature of Cognition

It is essential to acknowledge, before proceeding to the architectural critique, that the systems now collectively termed “neural networks” inherit their nomenclature from a biological analogy that — even at the time of its inception in the mid-twentieth century — captured only a partial slice of the relevant neurobiology. The intervening seventy years of neuroscience have rendered the analogy increasingly insufficient, and the commercial discourse around artificial intelligence has, in my view, systematically obscured rather than illuminated this insufficiency.

The dominant simplification underlying contemporary deep learning is what one might term the neuron-centric model of brain function. In this model, cognition is reducible, in principle, to the weighted summation of synaptic inputs across a sufficiently dense network of neurons, with learning consisting of the iterative adjustment of these weights through gradient-descent procedures. The model is mathematically elegant, computationally tractable, and — as Bengio, LeCun and Hinton (2015) demonstrated in their landmark Nature review — capable of remarkable feats of representation learning when scaled to industrial proportions. Its empirical successes are not in dispute. What is in dispute, or ought to be, is the extent to which the underlying neurobiological premise corresponds to how cognition actually operates in biological systems.

The glial cells of the central nervous system — astrocytes, oligodendrocytes, and microglia — were, for much of the twentieth century, dismissed as the supporting tissue of the nervous system, performing housekeeping functions of metabolic regulation and structural scaffolding while the neurons did the cognitive work. This characterisation has been substantially revised by the past three decades of neuroscience. Astrocytes in particular are now understood to participate actively in synaptic modulation, calcium-wave signalling, and the consolidation of memory traces, operating on timescales and through mechanisms that are categorically distinct from neuronal action potentials (Nedergaard, Ransom & Goldman, 2003; Verkhratsky & Butt, 2013). Fields (2009), in The Other Brain, synthesised the emerging consensus for a non-specialist audience: the brain is not principally a network of neurons embedded in passive tissue, but a co-computation between neurons and glia in which the glial substrate participates in functions previously assumed to be exclusively neuronal.

The point of this digression is not that contemporary artificial intelligence fails to be biological. The point, rather, is that the discourse around it routinely overstates how biological it is, and that this overstatement narrows the design space within which alternatives are considered. We have not constructed brains in silicon. We have constructed a stripped-down caricature of one component of one signalling pathway in one type of cell, scaled to industrial proportions and trained on the recorded outputs of human cognition. The architectural moves the dominant laboratories have made are descendants of an intellectual choice taken in the mid-twentieth century to model a particular abstraction — and abstractions, however fertile, foreclose as much as they enable.

II. The Conformist Machine: A Structural Critique of the Inductive Paradigm

The deeper critique is not biological but epistemological, and it concerns what the inductive paradigm of contemporary machine learning can and cannot, in principle, do. A system trained to model the distribution of a corpus learns, with whatever fidelity its scale permits, the shape of that distribution: its central tendencies, its regularities, the statistical texture of what has been written. This is an extraordinary achievement, and it is the source of the fluency that has so captured public and commercial imagination. But it carries a structural corollary that the fluency tends to conceal. A system that models a distribution is optimised toward the modal region of that distribution — toward consensus — and it has no representation of what is absent from the corpus, because absence leaves no trace in the data from which the distribution is learned.

This is the crux. Such a system cannot identify the abandoned hypothesis, the falsified route reconsidered, the contradiction that a discipline has quietly stopped attempting to resolve, the cross-domain analogy that no human has yet proposed because no human has read across the relevant boundaries. These artefacts cannot be retrieved by clever prompting or longer context windows. They occupy a region of epistemic space that the model has no representation of, because no representation of absence is present in the training data.

Marcus (2018) articulated a related critique with characteristic sharpness, arguing that deep learning is a pattern-extraction system rather than a hypothesis-generation system, and that the conflation of the two represents a category error of considerable consequence. Russell (2019), writing from a different vantage point on the question of AI safety and control, observed that the inability of present systems to reason about what they do not know is not merely a quirk to be debugged through future engineering effort but a foundational architectural property of the inductive paradigm itself. Bostrom (2014), several years earlier, had framed the deeper concern at the level of strategic posture: a system that optimises ferociously within a fixed objective function, without the capacity to interrogate the objective itself, is dangerous precisely in proportion to its competence.

What unites these critiques — and what the commercial discourse around contemporary AI has, for understandable economic reasons, been disinclined to confront — is the recognition that the systems now in production occupy the centre of the distribution of recorded human thought with extraordinary fluency, while remaining structurally incapable of articulating what lies beyond its edges. They are confident summarisers. They are not, in any meaningful sense, explorers. The distinction is not pedantic. It is, I shall argue, the determining variable in the geopolitical competition that is now beginning to take shape.

III. The Conscious-Incompetence Frontier: A Tractable Region of the Unknown

There is, however, a region of the unknown that is tractable — and the failure to recognise this constitutes, in my view, the central conceptual misstep of the present moment in artificial intelligence research. To delineate it requires the recovery of a framework that has, for nearly half a century, occupied a productive but somewhat marginal place in the pedagogical and strategic literature.

The four stages of competence model, generally attributed to Noel Burch and the Gordon Training International school in the 1970s, distinguishes four conditions of knowledge: unconscious incompetence (the things one does not know that one does not know), conscious incompetence (the things one knows that one does not know), conscious competence, and unconscious competence. Donald Rumsfeld, in his much-mocked February 2002 press briefing at the U.S. Department of Defense, articulated a parallel taxonomy in the language of “known knowns,” “known unknowns,” and “unknown unknowns” (U.S. Department of Defense, 2002). The framework was widely ridiculed at the time, but it has since proved durable in risk analysis, intelligence studies, and strategic planning. It is, moreover, deeper than the press conference suggests. Frank Knight (1921) had distinguished risk from genuine uncertainty eight decades earlier; Michael Polanyi (1966) had articulated the tacit dimension of knowledge that is, by definition, structurally inaccessible to formal articulation; Nassim Taleb (2007) had given the contemporary policy literature its sharpest formulation of consequential surprise as an irreducible feature of complex systems.

What the four-stage framework permits — and what, crucially, the inductive paradigm of contemporary AI has not engaged with — is the recognition that unconscious incompetence and conscious incompetence are categorically different problems. The first, by definition, cannot be searched. It is the territory of Kuhnian paradigm shifts before the shift, of conceptual frameworks that do not yet exist and therefore cannot yet be queried. The second, by contrast, is fully tractable. It is enumerable. It consists of the contradictions a field has explicitly identified and not resolved, the questions that have been asked but not answered, the techniques that have been proposed but not pursued, the hypotheses that have been raised and abandoned for reasons that may or may not be epistemic in nature. A system architected to map and probe the conscious-incompetence frontier of a knowledge domain would do something the present generation of artificial intelligence systems cannot. It would expand the surface area of the addressable unknown rather than the surface area of the already-known.

The intellectual lineage for this territory is long and, I would argue, has been insufficiently integrated into contemporary AI research. Stuart Kauffman (2000) named it the adjacent possible — the bounded but real space of innovations and configurations reachable from the current state of a complex system, but not yet realised. Charles Sanders Peirce (1931–1958) identified abduction as a third inferential mode, distinct from deduction and induction, and uniquely responsible for generating new explanatory hypotheses rather than confirming or extending existing ones. The Soviet engineer Genrich Altshuller (1984), working from a corpus of more than two hundred thousand patents, organised inventive logic around the contradictions a problem space contained, producing the methodology now known as TRIZ — a framework that, notably, plays a structural role in the methodological architecture of my own thesis as a horizontal analytical filter integrating evaluative criteria across multiple innovation ecosystems. Donald Stokes (1997) reframed basic research not as a pre-applied phase but as a use-inspired frontier — Pasteur’s Quadrant — whose internal structure is precisely the conscious-incompetence territory under discussion here. Joseph Schumpeter (1942), at considerably greater historical distance, had already located the engine of capitalist transformation in the discontinuities that disrupt equilibrium rather than the smooth extrapolations from it.

What unites these traditions is the recognition that knowledge advances through structured movement into the unrepresented. The current AI paradigm is engineered to do precisely the opposite — to consolidate and re-present what is already there, with ever greater fluency. It is, as a strategic instrument, a powerful but ultimately conformist tool.

IV. The Architecture of an Inverse Intelligence

The proposition that one might construct an artificial intelligence to “think the opposite of us” is, taken literally, a slogan rather than an engineering specification. Random or adversarial inversion of a model’s outputs would simply produce noise. The serious version of the proposition is structural rather than oppositional, and it requires the deliberate integration of a number of architectural primitives that exist, in scattered form, in the contemporary AI literature, but that have not been brought together under a unified objective function.

An Inverse Intelligence, as the term is used here, would be a system optimised not for the modal answer to a query but for surfacing the generative gaps in a knowledge domain. Its objective function would reward four classes of output, each of which corresponds to a distinct architectural primitive.

First, the identification of contradictions a field has stopped attempting to resolve, derived from the systematic analysis of citation patterns, abandonment rates of research lines, and the structure of unresolved disputes within a corpus. The TRIZ tradition (Altshuller, 1984) provides the conceptual basis; what is required is its computational generalisation.

Second, the recovery of lines of inquiry abandoned for non-epistemic reasons — funding cycles, ideological constraint, career pathways, classification regimes. This requires a model trained not only on what a field has produced but on the meta-structure of why particular productions were halted, redirected, or suppressed. The negative-space corpus, in other words, is as informative as the corpus itself.

Third, the discovery of cross-domain analogies absent from the literature because no human reader has spanned the relevant disciplinary boundaries. Schmidhuber’s (2010) formal theory of creativity and curiosity-driven learning gestures toward the technical primitives required; what is missing is the strategic deployment of these primitives against deliberately heterogeneous corpora.

Fourth, the detection of structural silences in a corpus — what is not written, given what has been written. This is, in many respects, the most demanding of the four primitives, requiring a model capable of probabilistic reasoning about expected literatures that have failed to materialise, and a hermeneutic capacity to distinguish epistemic silence from institutional silence.

None of these primitives is, in isolation, novel. What is novel is the proposal that they be integrated under an inversion objective, and that the resulting system be deployed as a strategic instrument embedded within an analytical framework capable of evaluating its outputs against the dual-formula architecture I describe below. It is also worth noting in passing that the most celebrated demonstrations of contemporary AI already contain incidental hints of what such a system might do. Move 37 in the second game of AlphaGo’s 2016 match against Lee Sedol — the move that human professionals initially called a mistake and subsequently recognised as a breakthrough — is, on the right reading, an inversion event (Silver et al., 2016). The system surfaced a strategy that lay outside the consensus distribution of recorded human play. It did so as a side-effect of self-play, not as a deliberate architectural objective. The proposal here is to make that side-effect the main effect.

V. The Ufimtsev–Lockheed Asymmetry: Historical Validation

The strongest case for the framework advanced here lies, I think, in the historical asymmetry it most cleanly explains. The case is well known to historians of aerospace innovation, less well known outside that subfield, and — when framed within the conscious-incompetence architecture proposed here — newly legible as the canonical demonstration of why an Inverse Intelligence capability matters.

In 1962, the Soviet physicist Pyotr Yakovlevich Ufimtsev published a paper titled Method of Edge Waves in the Physical Theory of Diffraction through the Soviet Radio publishing house in Moscow (Ufimtsev, 1962/1971). The paper was not, on its face, a strategic document. It offered a mathematical method for predicting how electromagnetic waves scatter off faceted geometries, extending the geometrical theory of diffraction in directions that, within the disciplinary frame of mathematical electromagnetism, were of moderate but not extraordinary interest. The paper was published, indexed, and freely available within the Soviet scientific literature. It was not classified, suppressed, or otherwise concealed.

The Soviet aerospace innovation ecosystem — at that time among the most institutionally sophisticated and resource-rich systems for translating scientific output into strategic capability anywhere in the world — did not see what it had. No Soviet stealth programme emerged from the Ufimtsev paper. No translation of its equations from predictive to prescriptive form was attempted within the relevant design bureaux. The paper sat, for more than a decade, as an item on the conscious-incompetence frontier of the Soviet system: its implications were, in principle, knowable, and yet they were not known.

The asymmetry was resolved from the other side. A mathematician at the Lockheed Skunk Works, Denys Overholser, encountered the Ufimtsev paper in English translation in the mid-1970s — the translation having been commissioned by the U.S. Air Force Foreign Technology Division and largely overlooked within the broader American defence-research community. Overholser, as Ben Rich (himself the director of the Skunk Works during the relevant period) records in his definitive memoir of the era, recognised what no Soviet engineer had: that Ufimtsev’s equations could be inverted. Where Ufimtsev had shown how to predict the radar scattering profile of arbitrary faceted geometries, Overholser saw that the same equations could be used to design geometries that minimised scattering in particular directions — that is, to construct an aircraft with a vanishingly small radar cross-section relative to its physical dimensions (Rich & Janos, 1994). The recognition catalysed the Have Blue technology demonstrator, which in turn catalysed the F-117 Nighthawk, which in turn anchored a doctrinal asymmetry in stealth aviation that has, in various refined forms, underwritten American aerospace dominance for four subsequent decades.

The instructive feature of the case, for present purposes, is the nature of the asymmetry. Both the Soviet Union and the United States had access to the Ufimtsev paper. The mathematical substrate — the equations themselves — was identical. No quantum of additional resources, no acceleration of existing R&D pipelines, no improvement in technological maturity scoring, no increase in resource investment within the framework of the Innovation Impact Formula would have closed the gap. The relevant frontier was not technological. It was epistemic, and it was institutional. One innovation ecosystem could see what was on its conscious-incompetence frontier; the other could not.

Epistemic status: the Ufimtsev–Lockheed history is documented — the 1962 publication and its 1971 U.S. Air Force translation (Ufimtsev), and Overholser’s recognition and its lineage to Have Blue and the F-117 (Rich & Janos, 1994). The reading of the case as an instance of a conscious-incompetence asymmetry, and as validation of the Inverse Intelligence framework, is the essay’s interpretive contribution rather than a claim of the historical sources.

This is, I would contend, the cleanest historical example of the problem an Inverse Intelligence is engineered to solve. A system whose explicit task is the systematic identification of Ufimtsev-class papers — open, unclassified contributions whose implications exceed their disciplinary reception, particularly across domain boundaries — is the operational instrument that could have closed the asymmetry from the Soviet side, or sustained it from the American side, depending on perspective. In the absence of such a system, both ecosystems were dependent on the contingent reading habits of individual mathematicians: a reliance on Overholser, on the Foreign Technology Division translator, on the unpredictable intersection of personal curiosity and institutional access. A state with a working Inverse Intelligence in the relevant period would not have been dependent on Overholser. It would have surfaced the Ufimtsev paper systematically, ranked alongside potentially dozens of other latent innovation candidates of similar character, evaluated according to a structured frontier metric of the sort described below.

The Ufimtsev case is not unique. It is, I would argue from the empirical record assembled across my doctoral research, a representative instance of a class of asymmetries whose detection has, until now, been the contingent product of attentive individual reading rather than systematic institutional capability. The Israeli development of low-cost unmanned aerial vehicles in the 1970s and 1980s, subsequently absorbed into U.S. military doctrine through the Pioneer programme, presents a structurally similar pattern. So, in a different register, does the trajectory of Soviet theoretical contributions to laser physics in the 1950s and 1960s, whose strategic implications were more rapidly translated into application by Western laboratories. The pattern, repeated across multiple cases, is consistent: the open scientific literature contains, at any given moment, a non-trivial number of contributions whose strategic and doctrinal implications exceed their disciplinary reception, and the geopolitical advantage accrues to the ecosystem capable of detecting and exploiting these contributions before its rivals.

VI. Integration with the Dual-Formula Framework: From Triad to Tetrad

The argument advanced to this point has proceeded principally as an exercise in epistemology and historical analysis. Its strategic significance, however, depends on the extent to which the proposed system can be integrated into the analytical architecture through which innovation strategies are evaluated and, in due course, adopted by states. It is to this integration that the present section turns, and in doing so it draws on the framework I have developed across the doctoral programme on aerospace innovation and geopolitical power.

The dual-formula framework I have proposed and applied across six historical case studies — the Innovation Impact Formula (IIF) and the Innovation Maturity Formula (IMF) — provides complementary quantitative tools for evaluating aerospace innovations within their geopolitical context. The IIF measures the immediate operational effectiveness of an innovation as a weighted function of urgency, adaptability, technological maturity, and resource investment. The IMF measures its longitudinal viability as a function of strategic weight and average time to fruition. Together, the two formulas disaggregate what is too often treated as a unitary phenomenon, separating tactical leverage from doctrinal absorption and enabling a structured comparative analysis across distinct innovation ecosystems and historical periods.

Both formulas, however, presuppose something that the framework itself does not provide: the identification of the innovation under evaluation as a candidate worth scoring. The IIF can rank stealth aviation as transformatively impactful; the IMF can demonstrate the durability of its doctrinal absorption. Neither formula addresses the prior and, I would argue, more fundamental question — the question that retrospectively haunts the Soviet aerospace system in the case discussed above: among the universe of latent innovations not yet recognised as candidates, which should we be looking at?

This is the conscious-incompetence frontier in its operational form, and it is precisely what an Inverse Intelligence system is engineered to populate. The integration I propose, accordingly, is not a bolt-on but an extension of the existing framework. A third formula — provisionally termed the Innovation Frontier Formula (IFF) — sits naturally upstream of the IIF and IMF, identifying candidates on the frontier before the existing instruments evaluate their impact and maturity. The IFF, as I am presently conceiving it, would evaluate latent innovation candidates against three principal variables: epistemic distance, the divergence between a candidate hypothesis and the modal consensus of its source field, normalised by the field’s internal heterogeneity; adjacent-possibility density, an operationalisation of Kauffman’s adjacent possible measuring the count and quality of latent innovations reachable from the candidate hypothesis within a bounded number of conceptual or technical moves; and cross-domain coherence, evaluating the formal isomorphism between the candidate hypothesis and structures in adjacent fields where its implications have not yet been recognised.

The full pipeline runs as follows. The IFF identifies candidate innovations on the conscious-incompetence frontier, populated by the outputs of an Inverse Intelligence system. Those candidates that survive frontier evaluation are then submitted to the IIF for assessment of immediate operational impact. Survivors of that evaluation are submitted to the IMF for longitudinal maturity assessment. The pipeline is sequential, the formulas are complementary, and the analytic architecture closes: from the unrecognised possible, through the operationally promising, to the doctrinally durable.

There is a further structural implication, which I think is among the more interesting consequences of the argument. The triadic model I refined in dialogue with subject-matter expert testimony — innovation, geopolitics, and operational influence, the third term capturing the field-level adaptation, improvisation, and repurposing observed empirically across both Cold War and Global War on Terror case studies — was, in its original form, capable of accommodating the phenomenon of initially dismissed concepts later proved superior once tested empirically only as a residual feature. The phenomenon was captured by the model but not directly explained by it.

The proposal of an Inverse Intelligence layer transforms this residual into a structural feature of the framework. The triad becomes a tetrad, with the epistemic frontier added as a fourth node: innovation, geopolitics, operational influence, and epistemic frontier. The fourth node closes a loop that the triadic model could not. Operational influence reveals which dismissed concepts proved superior in the field; the epistemic frontier identifies them before dismissal becomes the mechanism of discovery. The tetradic model, accordingly, accommodates the full innovation cycle: from latent possibility, through innovation candidate, through doctrinal absorption, through operational deployment, and back into the frontier as feedback for the next cycle.

This is not, I would argue, a cosmetic extension of the framework. It changes what the framework is for. The dual-formula model in its original form was a retrospective analytical instrument with predictive ambitions — capable of scoring innovations in their historical context and, by extension, evaluating proposed innovations against their expected impact and maturity profiles. The tetradic model with an integrated Inverse Intelligence capability is a prospective analytical instrument. It does not merely evaluate options; it expands what counts as an option. The distinction is, I think, considerable.

VII. Strategic Horizons: Toward 2050

The 2050 strategic foresight scenario developed in the closing chapter of my doctoral research posited that the primary arena of interstate competition will, over the coming quarter-century, shift from territorial dominance and even from technological capability narrowly construed, toward the management, protection, and acceleration of innovation ecosystems themselves. The argument advanced in the present essay suggests an additional and considerably sharper claim: that by mid-century the decisive variable in this competition will not be the size of the innovation ecosystem, the depth of its R&D pipelines, or even the agility of its operator–developer feedback loops. It will be the epistemic reach of the systems used to identify innovation candidates in the first place.

The reasoning is straightforward. A state that fields a working Inverse Intelligence does not merely accelerate its existing innovation pipeline; it changes what enters the pipeline at all. It converts the detection of Ufimtsev-class opportunities from a matter of contingent individual genius into a systematic institutional capability. Against a rival still dependent on the fluent but conformist systems now in production — systems that, however capable, are architected to think toward the centre of recorded thought rather than toward its edges — the asymmetry would be of the same character as the one that separated Lockheed from the Soviet design bureaux, but industrialised and continuous rather than singular and contingent.

This is the strategic case, and it is why the argument belongs in a defence-and-geopolitics register rather than a purely technical one. The dominant paradigm of contemporary artificial intelligence is the most sophisticated instrument of consensus ever constructed. It is, for that reason, the wrong instrument for the task that will matter most. The competition now beginning is for the first model architected to think against its own training distribution — to map, and ultimately to extend, the conscious-incompetence frontier of human knowledge. The state that wins that race does not get a better mirror. It gets, finally, a window. And the geopolitical consequences of that distinction will, I expect, define the architecture of strategic competition between now and mid-century.

References

Altshuller, G. S. (1984). Creativity as an Exact Science: The Theory of the Solution of Inventive Problems. Gordon and Breach.

Bengio, Y., LeCun, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. https://doi.org/10.1038/nature14539

Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.

Brose, C. (2020). The Kill Chain: Defending America in the Future of High-Tech Warfare. Hachette.

Bush, V. (1945). Science, the Endless Frontier: A Report to the President. United States Government Printing Office.

Christensen, C. M. (1997). The Innovator’s Dilemma: When New Technologies Cause Great Firms to Fail. Harvard Business School Press.

Fields, R. D. (2009). The Other Brain: From Dementia to Schizophrenia, How New Discoveries about the Brain Are Revolutionizing Medicine and Science. Simon & Schuster.

Kauffman, S. A. (2000). Investigations. Oxford University Press.

Knight, F. H. (1921). Risk, Uncertainty, and Profit. Houghton Mifflin.

Marcus, G. (2018). Deep Learning: A Critical Appraisal. arXiv. https://arxiv.org/abs/1801.00631

Mazzucato, M. (2013). The Entrepreneurial State: Debunking Public vs. Private Sector Myths. Anthem Press.

Nedergaard, M., Ransom, B., & Goldman, S. A. (2003). New roles for astrocytes: Redefining the functional architecture of the brain. Trends in Neurosciences, 26(10), 523–530. https://doi.org/10.1016/j.tins.2003.08.008

Peirce, C. S. (1931–1958). Collected Papers of Charles Sanders Peirce (Vols. 1–8; C. Hartshorne, P. Weiss, & A. W. Burks, Eds.). Harvard University Press.

Perez, C. (2002). Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages. Edward Elgar.

Polanyi, M. (1966). The Tacit Dimension. Doubleday.

Rich, B. R., & Janos, L. (1994). Skunk Works: A Personal Memoir of My Years at Lockheed. Little, Brown.

Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.

Schmidhuber, J. (2010). Formal theory of creativity, fun, and intrinsic motivation (1990–2010). IEEE Transactions on Autonomous Mental Development, 2(3), 230–247. https://doi.org/10.1109/TAMD.2010.2056368

Schumpeter, J. A. (1942). Capitalism, Socialism and Democracy. Harper & Brothers.

Silver, D., Huang, A., Maddison, C. J., et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. https://doi.org/10.1038/nature16961

Stokes, D. E. (1997). Pasteur’s Quadrant: Basic Science and Technological Innovation. Brookings Institution Press.

Taleb, N. N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House.

Ufimtsev, P. Y. (1971). Method of Edge Waves in the Physical Theory of Diffraction (U.S. Air Force Foreign Technology Division, Trans.). U.S. Air Force Systems Command. (Original work published 1962.)

U.S. Department of Defense. (2002, February 12). DoD News Briefing — Secretary Rumsfeld and Gen. Myers [Press briefing transcript].

Verkhratsky, A., & Butt, A. (2013). Glial Physiology and Pathophysiology. Wiley-Blackwell.