Introduction
Since Ancient Greek philosophy, reason has always had a special status. Within the
history of philosophy, reason has been described in two ways: on the one hand
epistemically and on the other instrumentally. The instrumental approach systematically
accomplishes the targets and the intents set out by a specific relation between thinking
and its surroundings — the enterprise of science, for instance. Instead, the epistemic
approach focuses more on mapping each belief, model, system to their respective
targets and on the efficiency of these cognitive and semantic maps. Explaining the
relations between reason, or thought, and being, or nature, has been considered one of
the tasks of reason.
Contemporary rationalist theories have emerged within the functionalist framework
inaugurated by American pragmatist Wilfrid Sellars and have tried to clarify those
relations without relapsing into a dogmatic understanding of the relation aforesaid.
For examples, Reza Negarestani1 emphasizes the autonomy of reason without
returning to a Cartesian rationalism). Although, the demarcation between reason and nature
— what Wilfrid Sellars defines respectively as space of reasons and the realm of
causes — is fundamental to answer certain issues, this should not make forget that
1 Following Hegel and Brandom, Negarestani sees reason as a protocol, a project which aims
for self-transformation [
23
].
rationality and its capacities supervenes on causal patterns recurring in nature.
Participating in the ‘game of giving and asking for reasons’ (GOGAR) is a key criterion to
distinguish between rational — or, sapient — entities and the rest of the entities which
inhabit the world [
1
]. In fact, the pragmatic independence of certain mental properties
and states is made more robust by the computational turn, which sees computational
neuroscience and artificial intelligence as contributing factors. Following the
developments in computational theory and in formal logic started from the beginning of the
20th Century, the pressures which are endured by various types of reason are rising.
An example of this kind might be offered by the enterprise behind Artificial General
Intelligence (AGI) — that is a kind of artificial intelligence capable of performing
human tasks, or capable of experiencing consciousness and awareness.
2
2.1
Contemporary Conjectures
Kant’s and Hegel’s Rekindling
In their way of responding and thinking how to articulate the relationship between
rule-governed rationality and pattern-governed behaviors, normative inferential skills
and bio-chemical structures, reason and nature, contemporary philosophers have been
rekindling two key figures of German Idealism, Kant and Hegel. Following
Enlightenment, Kant provides a vigorous theory of experience, which is justified for being
individual. In fact, particularized singular consciousness, i.e. ‘the transcendental unity
of apperception’ according to Kant, is constituted by a series of faculties, through
which perceptual, conceptual and agential capacities are considered as different.
Hegel is known for reinterpreting reason collectively and historically. He offers an
insightful idea in his theory of the Geist. Society — and more broadly culture — gives
rise to antinomies, schisms, divisions. Thus, a positive conception of antinomy should
be preserved.
Acquired from this theoretical background, reason distinguishes different
modalities of thought and practice and it recognizes them as part of a common framework.
Precisely, reason is able to organize the connection between identity and difference,
finite and infinite. It follows that reason not only represents the world but, as an agent,
it also progressively and effectively shapes it. Accordingly, the problem with Kant
rises when he absolutizes the limits of human knowledge by introducing the notion of
noumenon, which remains embroiled in causal determination. This removes the
systematicity Hegel tries to pursue and undermines the possibility of knowledge from
reason’s epistemological assumptions. Also, both Kant’s and Hegel’s conclusions
might lead to a naive idealist and thus anti-realist standpoint, which ontologizes the
demarcation between mind and nature, instead of rendering it pragmatical and
inferential.
2.2
Instrumentalism, Essentialism, and Its Counterarguments
Along with the rekindling of Kant and Hegel, the rationalist tradition has received
further underpinning by the practical results obtained by science [
2
]. The scientific
enterprise can historically be considered as antidogmatic thanks to its skills of
producing novel theories, hypotheses, and counterfactual arguments. In his short essay on
human knowledge, Karl Popper defines science by explaining two major theoretical
flaws which have been acquired during the course of history by the scientific
endeavour. The most recent one is instrumentalism. Popper does not completely dismiss
instrumentalism and, in fact, he tries to retain a feature of it. For this reason, his
conception of science holds an experimental instrumentalism, which conjoins a critical
discussion. Both are deployed to control scientific theories, opinions, and conjectures,
which are distinguished from pseudo-scientific theories for their aptitudes of being
falsified — rather than being supported — by any kind of evidence [
2
]. Thus, natural
sciences — and all applied sciences — do not belong to the realm of epistēmē or
technē, rather to the one of doxa, i.e. opinions and conjectures sensible to mistake and
trial [
2
]. What makes humans capable of being critical of their own creations is,
according to Popper, rational criticism and this specific human skill allows mind and
knowledge to transcend themselves and grow. The second view criticized by Popper
is the one of essentialism: as in the case of instrumentalism, he finds an obscurantist
feature which characterizes essentialism. This obscurantist aspect is the belief that
science has the ability to describe the ultimate essence of nature. It follows that
essentialism is confident that science is capable of assigning an undoubtful truth to its own
theories [
2
]. Popper argues that the essentialist point of view does not let thought
generate novel theories, issues, and conjectures.
2.3
Inferentialism
As in instrumentalism, Popper retains one feature of essentialism: “the scientist aims
at finding a true theory or description of the world, which shall also be an explanation
of the observable facts” [
2
]. From here, Popper departs to describe the third view,
which represents his standpoint. In short, the third view believes that while all efforts
in science are pointed towards a real description and explanation of the world,
scientists are aware that the results will never be definite — although, as Popper argues,
theories can be proved wrong with a high degree of certitude. Following Kant, he
argues that observable properties are falsifiable as data perceived through sense is
always elaborated with respect to a theoretical framework. Non-dispositional terms
have a relationship with experience, but the latter cannot verify the former. This leads
to an arbitrary decision on whether a scientific theory can be falsified or not by
nondispositional assertations. In other words, testing if an observable property is a true
falsifier of a conjecture becomes a trivial matter. Apart from being incoherent with his
falsification theory, Popper’s idea undermines his conception of truth based on
correspondence theory. The correspondence theory of truth is usually linked to a
metaphysical realism, which Popper defends. One among many oppositions of the
correspondence theory [
3
], inferentialism includes crucial aspects of all classic objections of that
theory of truth, i.e. coherentism, pluralism, pragmatism, and verificationism. Widely
understood, inferentialism is a result of rationalistic expressivism and rationalistic
pragmatism [
1
]; in fact, contemporary inferentialism is mostly elaborated by Robert
Brandom [
1
][
4-5
][
21
]. By shifting from representationalism to inferentialism (hence,
privileging inference over reference and denotation), Brandom draws inspiration from
Frege, Dummett, and Sellars, explaining meaning by way of the inferential
connections between linguistic expressions [
6
]. The pragmatic approach of languages is a
reformulation of Sellars’ doctrine ‘meaning-as-use’. Therefore, inferentialism
understands semantic statements as assertations which predicate an inferential relation
between a term and an object or a property of such object: speaking does not assign any
property to reality, rather it becomes a matter of establishing properties and roles.
Here, within the general inferentialist context, James Trafford goes beyond
Popper’s notion of truth. In his Meaning in Dialogue [
6
], his main aim is to show that
meaning is a result of a dynamic interaction between agents, who act in a
nonlogically pre-determined space, as logical rules emerge through these interactions.
While he is aware that Brandom’s inferentialism eases agents from questioning their
relationship with mind-independent entities (as inferentialism shifts the issue to the
establishment of normative properties [
6
]), Trafford also claims that standard
approaches to inferentialism — usually based on monotonic, static, rule-governed
methods [
6
] — fail to meet his thesis; in fact, they dismiss the dialogical and interactive
role of proofs, refutations, assertions, and denials. Trafford rejects Popper’s belief:
counterexamples for a theorem’s validity cannot necessarily be found in the future [
6
]
as this would mean dismissing any symmetry between proof and a conclusive
refutation of a theorem. Popper’s asymmetry is not accepted because it would allow a set of
logical rules to casually determine interactions [
6
]. Trafford, instead, defends the
dialogical interaction between proof and refutation free of normative constraints. The
alternative offered is a constructive, defeasible, and pluralist approach to logics and
inferentialism.
3
3.1
Contextualising Functionalism
The Origins of the Functionalist Approach
Functionalism is a theory about the constitution of mental states. According to
functionalists, mental states are defined by what they do rather than by their internal
constitution. Historically, functionalism emerged following the failure of behaviourism
and the cognitive revolution in psychology. One of the prominent psychologists who
describes stimulus-responsive (S-R) behaviourism’s failure is George Mandler. In his
paper Origins of The Cognitive (R)evolution [
7
], Mandler stresses that the gradual
shift from a behaviourist psychology to a structuralist, cognitive, and functionalist one
occurred between 1955 and 1966. His main opposition to behaviourism is that classic
S- R theory and radical behaviourism fails to explain issues around human thinking
and its agency and memory because of the undifferentiation between humans and
nonhuman animals [
7
]. In fact, even Burrhus F. Skinner’s functionalist behaviourism
— which was aware that to illustrate a being as a functional system thereby entails
characterizing it not in terms of basic regularities (defined as reliable differential
responsive dispositions by Brandom [
21
]) reducible to patterns of S-R behaviour—
lacks an analysis on specifically human functions [
7
].
Functionalism as a research strategy has another point of divergency with classic
behaviourism, i.e. the multiple realizability thesis, which is one of the most important
arguments for functionalism. By describing mental states in non-mental language —
that is, logico-mathematical language — functionalism meets one of the requirements
of behaviourism [
8
]. However, functionalism defines mental states in non-mental
terms by quantifying over realizations of mental states [
8
]. Instead, behaviourism
keeps its focus on environmental factors and only quantifies over external
conditioning. Functionalism holds the ideas that functional states can be realized in different
ways; vice versa, one physical state — e.g. pain — can realize several functional
states in different devices [
8
]. Apart from clarifying the relation between mental states
and their explanations in terms of logical and conceptual properties, functionalism
also elucidates on how to distinguish sensible data from semantic inferences with
regard to mental states’ functional roles and relations.
3.2
Computational Functionalism
Alongside its recent history, functionalism has been understood in different ways
within a philosophical outlook. The two main sources of it are the empirical
computational theory of mind, chiefly explained by Hilary Putnam, and the functionalist
theory of meaning sketched by Ludwig Wittgenstein and elaborated by Sellars. Machine
state functionalism [
9
] corresponds to an early elaboration of a computationalist and
functionalist theory of mind. Indeed, Putnam’s theory is among the firsts
computational theories of mind (CTMs) — today known as classic computational theory of
mind (CCTM). According to Putnam’s theory, mental states operate similarly to a
Turing machines and thus they share the same properties, meaning that a set of
instructions function as an algorithmic rule-set in relation to certain system parameters
to determine a certain mental or machine state. It is important to remark that the
resemblance between the two models has never been understood as an integral blending
because abstract Turing machines are deterministic, and they have sequential
processes and an infinite non-addressable memory. Whilst, minds are more stochastic, they
can execute parallel computations by generating different types of outcomes from
inputs — while inputs and outputs in a Turing Machine are always symbolic — and
they have a finite addressable memory.
Surprisingly, Putnam himself became a radical adversary [
10
] of his theory and,
more generally, of every kind computational functionalism years later. Through
Gödel’s theorems of incompleteness, he develops four arguments to confute
computational functionalism. All his theses are refuted by Jeff Buechner [
11
], who, by
preserving the importance of Putnam’s theory in cognitive sciences and inductive
reasoning, provides a history of the applications of Gödel’s theorems and the different
degrees of relevance of computationalism within functionalism. In fact, in
Representation and Reality [
10
], Putnam argues that inductive reasoning cannot be formalized,
and he proves this with the incompleteness theorems — which have now gained an
epistemic value [
11
]. It follows that there cannot be any computational description of
the human mind or of general intelligence as both the computational and the
noncomputational inductive mind cannot prove mathematics’ consistency. All methods of
inquiry are thus susceptible to Gödel’s theorems. Yet, Buechner questions why
humans, who, according for Putnam, have no computational descriptions, can succeed in
computational tasks [
11
]. On the one hand, he believes no convincing reasons to
dismiss computationalism are offered by Putnam. On the other hand, what it is
interesting to highlight from Putnam’s R&R is the view that a general intelligence does not
consist — only — in an inductive account of reasoning, but — also — in an abductive
and deductive version. Inductive reasoning can only be dismissed if it was proposed
as the only mode of mind’s functioning. Rather, if the three forms of reasoning are all
included in the debate around general intelligence’s automatization, the inductive type
of thinking can still play a crucial role.
3.3
Pragmatic Functionalism
Putnam takes part in solving the riddle of induction by extending his critique to a
description of mind — i.e. inductive reasoning cannot be formalized by a human
mind. According to Buechner, he fails encountering Kripke-Wittgenstein
rulefollowing paradox [
12
]. Far from following Kripke’s interpretation of the paradox,
John McDowell [
14
] sees Wittgenstein’s dismissal of his paradox [
13
] not as naïve,
rather it is coherent in respect to his view, which conflates understandings with
interpretations. Following a rule derives from the inculcation of the rule into a cultural
practice. One learns how to use ‘plus’ because it was instilled into the practice of
adding. This usage-based theory of language appearing in Wittgenstein’s philosophy
influences Sellars’ version of pragmatic functionalism. Sellars’ researches on thought
represented a crucial shifting in analytic philosophy2. In contrast to the scientific
image of humans, his manifest image of humans — which consists in the
quasitranscendental property of reasoning instantiable in other systems than neurobiological ones
— rekindles the Kantian project and other problematics regarding reason’s
infrastructure. In Empiricism and the Philosophy of Mind [
15
], Sellars defends a kantian theory
of mind by refuting any Cartesian view of mind. His approach moves towards a
naturalism, which does not neglect the normative dimension of mind. Accordingly, Sellars
believes that psychological concepts should not be considered less relevant than
theoretical concepts in the sciences: mental concepts have the same importance of any
scientific concept within naturalism.
The implication of Sellars’ top-down approach [
16
] is vital: mental states are not
expressed in specifically human terms, rather they are articulated within a non-human
framework as they correspond to functional tokens. Computational descriptions
become possible because conceptual behaviours are now understood as transmitters of
tokens which maintain inferential relations with other functional states. As long as a
computational system becomes capable of conveying semantic content by
instantiating interactive, dialogical processes between itself and other computational agents, it
can enter Sellars’ interactive and inferential space of reasons. Inferences are
fundamental for the GOGAR as they establish normative properties between assessments.
2 Under the guise of Kant’s philosophy, Sellars’ theories allowed to surpass the Humean
deadlock in which analytic philosophy was stuck [
1
].
Within this framework, reason perceptively responds to an input, and elaborates an
action in virtue of a protocol which aims to organize the cognitive and behavioural
resources of a system in order to achieve a precise task. Across the space of reasons,
terms can change their status from observational to theoretical, or vice versa, and
multiple states can be implemented, realized, and instantiated in different
technological devices in different ways. Furthermore, epistemic and ethical attitudes can be
found accompanying reason. The responsibility in committing to a claim and the
normative conceptuality which surround reasons allow minds — or, computational
systems — to construct knowledge. It is thus necessary to introduce theoretical
underpinnings to explain and track the divergence between the scientific image and the
manifest image in order to describe how cognitive properties which yield sapience can
be instantiated in machines.
4
4.1
An Engineering View of AI
From Correlation to Counterfactual Judgments
Judea Pearl is one the few machine learning (ML) researchers who have recently tried
to go beyond the current paradigm ubiquitous in this field, i.e. deep learning. During
his academic years, Pearl focused on probabilistic inference [
17-20
]. His aim is to
challenge standard probability calculations — i.e. reasoning by association —
embedded in deep learning. In fact, he is one of the pioneers of Bayesian neural networks
(BNNs), which, according to him, allowed machines “to think probabilistically” [
17
].
He believes that current machine learning techniques pay too much attention to
uncertainty, while they are not concerned with causal reasoning — that is, thinking cause
and effects. The pure statistics applied in ANNs with the curve fitting method shows
that their form of predicting, a correlationist approach to AI cannot be accounted as a
form of intelligence. This general idea is what leads Pearl to describe the ascension
towards the ‘ladder of causation’. Thus, he reconstructs the history of probability and
causality starting from Thomas Bayes and David Hume.
During the 18th Century, Bayes gave an analysis of ‘inverse probability’, which is
the opposite of forward probability. Bayes was interested in finding the probabilities
of two events: the hypothesis and the evidence. Given that one knows some of the
prior conditions, Bayes calculated what the probability for a certain event to occur is.
He eventually came out with a general solution, which is known as Bayes’ rule: by
estimating the conditional probability in one direction, it is possible to mathematically
derive the conditional probability in the opposite direction., which is less reliable than
the former [
17
]. Thus, Bayes’ rule expresses how a degree of belief, accounted as a
probability, should rationally be updated to justify for accessibility of associated
evidence. In other words, Bayes’ rule is needed to update beliefs in a certain hypothesis
as evidence can never be fully certain. In this fashion, Pearl and other scholars
imagined a hierarchical architecture for ANNs — i.e. BNNs — which could pass
information on higher orders through conditional probabilities and likelihood ratios. BNNs
are gifted with building blocks similar to causal networks [
17
] in order to imitate the
rational process, which understand cause-effect relationships.
As it is easy to imagine, Bayesian networks highlight the crucial connection
between causes and probabilities, which helps estimating the probability of a pattern of
dependencies in the observable data, given a pattern of network in the building block
[
17
]. Pearl’s desiderata involves eliminating do-operators from Bayes’ rule to reduce
the estimation to a less complex calculus. In other words, reducing the ‘do’ quantity
to a manipulation of the observable data [
17
]. Do-operators are introduced by Pearl
[
18
] to resolve epistemic problems in probabilistic causation. They are expressed as
P(effect|do(cause)) in contrast to classic Bayes’ rule expressed as: P(effect|cause)
[
17
]. While the latter indicates a passive observation, the former shows an external
intervention. Mere observational cases cannot establish any cause-effect property
between two cases. This confirms the most famous phrase in statistics, i.e. correlation
does not imply causation. However, conscious manipulation (the do-operators) may
estimate causal effects from observational experiments. Pearl argues that Bayesian
networks can only perform the do-calculus, which generates the reduction — i.e.
replace an experiment with a mathematical model — he aims to achieve as they can
preserve the meaning of a probabilistic expression [
19
]. Furthermore, completeness is
a property of the do-calculus axiomatic system, meaning that that decision problems
are controllable [
19
]. If decision problems are manageable, it follows that an
algorithm can decide whether there is a solution to an expression or not in a finite and fast
time, i.e. in polynomial time [
19
]. Thus, Pearl3 shows how to automate statistical
expressions strongly related to causation with mathematics.
However, the final element which makes Pearl climb the ‘ladder of causation’ is
not the algorithmic version of the do-calculus. The quintessential element for causal
reasoning is the counterfactuals as it helps producing more efficient and precise
statements on several issues [
19
]. Imagining different outcomes becomes possible by
adding or removing certain causes from an observable phenomenon. Therefore,
‘what-if’ form facilitates defining a cause as necessary, sufficient, or
necessarysufficient. According to Pearl, Hume was the first philosopher to include
counterfactuals in his account of causality. Hume suggests two opposite definitions for causality.
On the one hand, he defines causality as a mere observable pattern regularity which
occurs in nature. On the other hand, he introduces the notion of counterfactual
judgment and thus he establishes a cause-effect relationship. Pearl retains the latter
definition. Counterfactual judgments consist in imagining alternative worlds where a
different action would have produced another outcome. These structural models are the
short-cut which allows to simulate and compress infinite imagined worlds, to compute
the closest to our reality and, which ultimately allows causal reasoning.
Being capable of extrapolating patterns within certain degrees of uncertainty is not
enough for Pearl. He is aware that Bayesian networks are not capable of establishing
causal properties [
19
]. Instead of focusing on opaque systems — e.g. CNNs — and
mere detection of patterns within a data set, Pearl is still trying to instantiate
consciousness, awareness, and intentional states in technical devices. These metaphysical
elements constitute the notion of agency. Therefore, next steps in AI should aim to
provide a software model for its agency. This hard engineering problem is pursuable
3
With the help of other scholars and his students.
for two reasons. Following Sellars, mental properties are virtually instantiable in
technical device. Following Pearl, algorithms for causal and counterfactual issues
already exist [
20
]. Accordingly, two further questions follow: what is the hierarchy
behind cognitive skills identifiable in sapient agents? Which theoretical framework is
necessary to instantiate mental properties in technology? An overview and a critique
of Brandom’s conceptual hierarchy; an illustration of a novel functionalist theory, i.e.
the engineering approach to functionalism.
4.2
Brandom’s Conceptual Hierarchy
Pearl seems suggesting an engineering approach to computational functionalism.
Before elaborating this idea, it is necessary to highlight a certain hierarchy among
non-sapient and sapient cognitive skills. This hierarchy removes any kind of naivety
from theories aimed to level non-sapient and sapient skills, which — as it has
described so far — only omits fundamental characters of what is understood as
consciousness, agency, and reason. The peculiarity of functionalism is that it does not
treat consciousness, agency, and reason as something exclusively human meaning that
implementing sapient cognitive skills in technology is possible — which is
scientifically (and partly) confirmed by Pearl. The conceptual hierarchy developed by
Brandom is a ladder to rational inferentialism, which, by virtue of the theoretical
assumptions illustrated so far, is more robust and less vague than Pearl’s ladder to causation
and is coherent with Sellars’ functionalism [
15-16
].
In How Analytic Philosophy Has Failed Cognitive Science [
21
], Brandom
describes how analytic philosophy failed to investigate and strengthen Frege’s theories4
present in his Begriffsschrift which gave birth to analytic philosophy itself. Following
this failure, analytic philosophy has not convincingly clarified issues around the
principles around mind, reason, and thought. Thus, analytic philosophy has also failed
cognitive sciences [
21
] which, in turn, have made several important progresses in
explaining how to empirically and experimentally perform cognition. Philosophy,
instead, must move in another direction according to Brandom. Following Sellars,
Brandom argues that philosophy should ask normative questions around conceptual,
sapient, discursive practices of sapient entities. To pursue such enterprise,
philosophers should analyse the structural cognitive hierarchy within concept-use practices
[
21
]. Repurposing this ladder towards the achievement of causal inferential modalities
of AGI means refuting several accounts of experimentalism and non-pragmatic
heuristics. The scientific enterprise is a crucial epistemic practice in the neo-rationalist
movement and Brandom’s ladder confirms the developments made in the first
chapter, where, following Kant and Hegel’s contemporary revitalizations, debate within
philosophy of science were introduced.
Brandom’s hierarchy involves four basic steps: labelling-classification — the level
of bare reliable differential responsive dispositions; describing — the level of rational
classification; formal inference — the level of logical classification; complex
predica4 I.e. the mathematical characterization of semantics and conceptual content, and so of the
structure of sapience [
1
].
tion —the level of invariance under substitution. Conceptual, rational classification is
an attribute of sentient entities and it encompasses an abstract clustering of something
particular into a general term, which expresses a similarity emerging among
particularities. Applying a concept to something means describing it, which is informative as
it implies that something follows the description. The semantic import arises when
something is inferred from a description [
21
]. Instead, labelling an object has no
informative content and thus a label does not have any conceptual consequence. In fact,
the last distinction Brandom stresses is between simple and complex predicates.
While a simple predicate occurs as a component in a sentence, a complex one include
any multi-place simple predicate. According to Brandom, through Frege’s logical
notation it is possible to form complex predicates, and logically codify the inferences
that articulate its descriptive content. In fact, Frege’s notation allows to invariantly
substitute singular terms with variables and quantify these variables in order to form
complex predicates. Thus, the final step of Brandom’s hierarchy, i.e. the analytical
concept formation [
21
] described with Frege’s notation, accounts for concepts formed
at the highest cognitive stage, which represent the most expressive concepts of the
whole ladder.
Making sense of the world, organizing it, and changing it by optimizing one’s own
models are sapient capacities, which occur through inferential semantics, rational
pragmatics, and systematic theoretical trajectories. The challenge of AGI should be
slanted towards the development of these inferential skills. Although Brandom’s
ladder is clearly useful for the project behind AGI, an engineering approach to
functionalism is still needed to account for the conditions which determine the way sentient
and sapient skills could be implemented in technical systems.
4.3
An Engineering Approach to Functionalism
Among all the functionalist distinctions, the engineering approach is one that asks
about how we might be able to replicate and simulate mind and its particular states in
other, inorganic media. This question is not only about how we replicate the human,
but also about how we can engage in processes of synthetic production by virtue of
which different media can serve as a prosthetic outsourcing of sentient and sapient
abilities, e.g. new kinds of perceptual experience, new forms of cognitive heuristics,
systems that have non-textual, non-discursive or vocal languages but that
communicate in other ways, etc. It can be argued that modelling, prediction, pattern
recognition, memory, language processing and probability are cognitive procedures that can
be instantiated in media with new shapes and methods5. However, imagining artificial
entities as epistemic agents is not fiction. Imagining them as conceptual agents might
still be. This means that even though technical devices influences human’s way of
reasoning, act, think, they cannot perform the high-order conceptual behaviours yet.
Engineering functional properties means avoiding any trivial levelling of sapience,
5 Far from endorsing any panpsychist view of mind, — both Brandom and Pearl agree that it
is necessary to state that state-of-art of AI is stuck at the level of non-conceptual labelling,
i.e. pattern detection through correlational statistical means.
reason to nonconceptual activities. And, an engineering approach to functionalism
never starts with the presupposition of a vague superiority of bio-chemical human
structures over other synthetic systems by virtue of the well-known multi-realizability
argument.
Michael Weisberg [
22
] is instrumental in highlighting how the models used by
engineers do not include those assumptions. This shows how local applications of
models have always global unpredictable consequences in the process of engineering. For
Negarestani [
23
], the pragmatic and epistemic method which engineers exploit is a
way to see the project behind constructing intelligence as a problem of scaling
different models, which interact as the explanandum and explanans of them. Thus, every
theorist can access the epistemic metatheoretical assumptions as Weisberg
pragmatically argues that the core of scientific evaluation and judgments consist in the analysis
of the relations between models and reality. These relations are complex and plural,
rather than merely ineffable and inaccessible. Interestingly, one of the consequences
of the engineering approach is not a reduction of reason to a mere experimental
heuristic, but a coherent conceptual image of reality globally valid and in constant
development. Models in engineering are continually weighing their assumptions, tuning
their theories on different scales, calibrating their scope between a global image and a
local phenomenon as both universal and narrow perspective are always needed. An
AGI, a project about the instantiation of conceptual capacities typical of sapient
entities in technical devices, should thus move in this direction, i.e. towards an updated
dialectics between different degree of scopes by understanding the peculiarity of each
level, the multi-level relations, the proper method of explanation, description, and
prediction of the either global or local stage of representation, and the inter-play of the
many dynamic, defeasible, non-monotonic inferential logics.
5