Foundations of Artificial Intelligence Frameworks: Notion and Limits of AGI

1 Understanding Artificial Intelligence

The topic of artificial intelligence has been debated since the long form of intellectual pursuit, started with, not as comprehensive starting point as possible, Turing (1950); McCarthy et al. (1955). The conversation following such cultivated into different voices of definition and opposition via problems, such as McCarthy (1987) on the definition of strong and weak AI, Searle (1980b) definition of Chinese Room Argument, Russell and Norvig (2010) definition of intelligence on basis of agents action, Penrose (1989b) argument against the philosophy of computational intelligence, Floridi (2004)’s ethics. More can be found on Stanford Encyclopedia of Philosophy (2018) and further sources, but it is sufficed to say that the topic has received no less of contributions thereof to itself. However, it is reasonable to see that such argument and debate is often very shallow and hypothetical.

Inside the term artificial intelligence, there is the word intelligence. A normal person will tell you that they are intelligent. But it just so happens that this notion of qualification is harder to define when one participates in the active action of finding it. So, what is it? This is the question we should take in.

That said, this question is very much arduous in its public view, and distastes of the technical crowd. just as Øygarden (2019) mentioned, for philosophers whom are assumed to be interested in such endeavour, the topic of intelligence has traditionally appeared of a less interesting concept than consciousness. Pardon their minds and view, the philosophy of mind and body seems to attract more on the front of abstract thought, than something considered to be mechanical by thinking, and attribute no more and no less to the notion of being human as it is, with the soul and body being the main question. Developments of AI have caused new philosophical interest in the concept of intelligence, though seldom appropriately decoupled from its closely related phenomenon consciousness. Even in the field of AI itself, there are the avoidances of defining intelligence, though there have been no shortage of finding one in the middle of the forest. It is said that, for someone to work in the field of artificial intelligence, it would be wise to couple oneself with a definition of AI on him/herself, rather than not. Such is to say as to fix a philosophical standpoint before working in the field, which both contributes to the enormous amount of opinionated definition, but also the rigid framework on which artificial intelligence is considered.

As such, there exists no satisfying definition of artificial intelligence beyond the notion of artificial, of which is still dubiously believed upon. However, to fully capture the notion of AGI, we need the notion of AI on it.

1.1 Understanding artificial

What separated, of the artificial and the ‘natural product’ by definition? Below, we present the table on such terms of the concept ‘being artificial’ taken from the most basic of the knowledge, etymology and definition available in large, of Merriam-Webster (2025); Cambridge Dictionary (2025); Oxford Learner’s Dictionaries (2025); Dictionary.com (2025); Justia Legal Dictionary (2025); Merriam-Webster Etymology (2025); Bianchini (2021); NASA (2023); IBM (2024); Legg and Hutter (2007b); Goertzel (2014) and so on so forth. Indeed, for as long as the field artificial intelligence is formed formally of 1956, in the Dartmouth workshop, those are what we understand of the concept of artificial intelligence, as to be normalized of into triviality. In the process of making something artificial, one must then have to reproduce what they considered natural, in certain perspective — not man-made — into a man-made form, with certain criteria. This is discussed in Simon (1969); Haugeland (1985); Boden (1987); Boden (1996); Boden and Edmonds (2019); Onyeukaziri (2022); Boden (1990), in particular, about natural and artificial concept.

Table 1: Comparative definitions of artificial across different domains

In all, the definition of artificial can be considered diluting, since essences and the term is usually wholly considered in different voices and perspectives. For example, one might mistakenly classify biological to being natural --- certainly a farm can be just as biological as a plain field without intervention, but not natural as it is[^1]. On another matter, every object that is man-made still follows the law of physics, they obey the law of which is itself an expression of how the universe works, but such artifacts are called artificial, not natural, even though a rock still obeys such rulesets. We can then see that to equate something with the property of being artificial, requires more than just putting on its label, for the term natural itself is very hard to grasp. The science of artificial itself is now being advocated, much to Simon (1969) idea that a lot of the current world is now artificial in perspective. However, one can propose such dilemma by equating artificial with something entirely: process. Simply put, we define natural as without intervention. By our scientific fact, we know that the world changes and evolves itself, develops and mutate by time, of the second law of thermodynamics. Then, we can define something being artificial as a thing created, from a process outside such evolution, of which then subjectively, means created by human. To be created by human also means to conform to the notion of interpretation, encoding, of which is then represented in certain ways or form. While philosophical, it is trivial to see that without interpretation, an object itself will remain as plain of the object itself, without purposes, without triviality. Just as a neural network without the interpretation to be a neural network is useless, and just as mathematics is useless without the interpretation of the subject and objects of consideration. Thereby, we then accept such provisional overview as the definition of artificial, from the subjective view of being created by human, hence constructible, and is created of purposes, its process and operations can be given meaning and representation.

The definition has a drawback, though, that is, if we later on discover an alien species with the capability to construct such artificial intelligence, would we not call it artificial intelligence, because it is not made by human? We need a definition that generalize to certain notion of intended creation, for in such event if we ever discover alien lifeform, then their ”AI” would not be argued against such to be not artificial because of the word’s basic meaning related to human, and hence can be considered a stipulative definition in its stead. Doing such, requires the notion of intention, and the notion of reason.

Definition 1.1 (Artificial)**.**

A subject $A$ is artificial, if it is created by an object $B$, with intention and tasks $p$, constructed in a way of which the logic system of $B$ can interpret such object, the representation of $B$ can simulate and construct $A$, and $A$’s operation lies within such domain of interpretation and logic of $B$.

The notion of artificial will always be troublesome, as it is based on the ground that there exists certain subject of the current universe, $B$, that can facilitate and create another object $A$ of various means, with interventions and interferences from one’s natural evolution. While not satisfactory, it will be our provisional definition on the topic of artificial. In a sense, however, it is wise to note that the definition and treatment of the term artificial is purely subjective, per subject in question. The fallacy of such provisional definition comes in the very simple thought experiment. If ants are to create its own fungus farm, would we call that fungus farm artificial, or natural? Indeed, one can attribute that to the fact that the being, observer is attributing such terms and meaning to the structure that is interpreted as fungus farm, while in fact in natural it is just a natural behaviour of the ant colony. However, how far will such interpretation stretch, and edge cases of which it can handle is unknown. Where does nature end, and where artifice begin?

Of such basis, of the artificial platform in which we as subject $B$ create, artificial intelligence theory then posit that the notion of intelligence can be represented, interpreted, and constructed using a system called computer. Though variations differ of which what is considered computation, how is symbolic manipulation considered, this theory is the philosophy of Computational Intelligence, or Computationalism, of which formed the basis of artificial intelligence research of date. Supporters of the theory include as earliest being Putnam (1988); Fodor (1975); Churchland (1986); Dennett (1978); Dennett (1991), and of the 21st century, Scheutz (2002); Dodig-Crnkovic (2012); Gauvrit et al. (2015) and Copeland and Proudfoot (2018). This theory then posits that we can represent intelligence, of the arbitrary definition that it is understood upon, using the language of representation of computer, using numbers and computational processes as for simulation, using algorithms to simulate thinking processes, and else. Of course, there exists pushback against such notion, of which creation of artificial intelligence is not totally configurable using computers, notably Searle (1980a); Searle (1992); Penrose (1989a); Penrose (1994); Nagel (2012); Müller (2025); LaForte (1998). To answer if the theory is false or not, and what is the implication of such theory, we have to shift to understanding the second term of the word — intelligence. Furthermore, we also have to realize what is said as a computational structure, and as the artificial construct that we created of the computational framework, how and what will constitute the creation of artificial intelligence in such specific representation.

1.2 Understanding intelligence

Inside the term artificial intelligence, there is the word intelligence. A normal person will tell you that they are intelligent. But it just so happens that this notion of qualification is harder to define when one participates in the active action of finding it. So, what is it? This is the question we should take in.

We start this section with a series of historical accounts. Shane, Marcus (2007)’s paper A collection of definitions of intelligence,Legg and Hutter (2007a), and Masahiro’s (2023) Descartes and Artificial Intelligence[^2] might be a great place to start this, since they provide a non-trivial amount of definitions and attempts already there, which serve us more as exhibition for observant in this section and beginning.

R. J. Sternberg …I prefer to refer to it as ’successful intelligence.’ And the reason is that the emphasis is on the use of your intelligence to achieve success in your life. So I define it as your skill in achieving whatever it is you want to attain in your life within your sociocultural context — meaning that people have different goals for themselves, and for some it’s to get very good grades in school and to do well on tests, and for others it might be to become a very good basketball player or actress or musician. D. K. Simonton …certain set of cognitive capacities that enable an individual to adapt and thrive in any given environment they find themselves in, and those cognitive capacities include things like memory and retrieval, and problem-solving and so forth. There’s a cluster of cognitive abilities that lead to successful adaptation to a wide range of environments. H. Nakashima Intelligence is the ability to process information properly in a complex environment. The criteria of properness are not predefined and hence not available beforehand. They are acquired as a result of the information processing. P. Voss …the essential, domain-independent skills necessary for acquiring a wide range of domain-specific knowledge - the ability to learn anything. Achieving this with ’artificial general intelligence’ (AGI) requires a highly adaptive, general-purpose system that can autonomously acquire an extremely wide range of specific knowledge and skills and can improve its own cognitive ability through self-directed learning. Jensen, Huarte, Dearborn …the ability to learn, the ability to understand, either principles, truths, facts, or common sense, to profit from experiences; the ability to comprehend, or the capacity to reason. A. Anastasi Intelligent is functionally of multiple components combined. J. Peterson …a bunch of stimuli. Humphreys …the resultant of the process of acquiring, storing in memory, retrieving, combining, comparing, and using in new contexts information and conceptual skills.

Out of those definitions, there are two kinds of defining the notion of intelligence, we call it the top-down and the ground-up approach. The top-down line of thought demonstrate, most of the time conjectures, the existence of intelligence as a whole, without finding the actual shell that contains it. If intelligence is general, then their implementation follows, but to a sufficient degree, it can be achieved everywhere. It guarantees, partially, of certain school of thoughts the generalizability of intelligence as the ground base to re-create such, which is characterized, often, by current machine learning discipline. This approach beside from guarantees such existence, also has the capability to ‘test’ a subject of being, ‘intelligent’.

This is done by setting up agenda and criteria, of which the current theory serves as more of a black box for the actual ‘machine’ that contain it, but enough exhibitions fitting those criteria for intelligent. Fortunately, this also sets certain criteria for artificial intelligence to be specified so in the name. The Turing test, which posit different observable properties to be examined, and the Gödel’s argument is one of such example in this line of thoughts, theorized by J. R. Lucas (1961), Penrose (1994, 1989), and Benacerraf (1967), similar to the Chinese Room Argument (Searle, 1980). Coincidentally, the notion of computationalism is also formed out of this approach. In a sense, it works as the following definition.

Definition 1.2 (Intelligence, top-down approach)**.**

We say that we observe intelligence in any given circumstances, of any arbitrary object regardless of structure, if it exhibits observable behaviours to the environment, the surrounding, the interested space such that can be clarified, and identified, to the nearest high-intelligent specimen (human), to certain degree of operational arbitrariness, of its own activities, properties, and functions. In such case, intelligence is defined per speculative reference point (human) and of criteria that fits the such point (human) model of intelligent.

The definition in the top-down sense is then entirely subjective from the human perspective.

The ground-up approach of defining intelligence is simply the polar opposite: Instead of defining intelligence by criteria, they create machines or models that have intelligence seems to be the emergence behaviour from those model. That is to say, they define intelligence by not defining it but constructing it. Though, this type of approach still requires the intuitive feeling of intelligence to figure out or identify such emerging signs of a growing construct, but it is more or less general, as it does not depend on certain opinion, or fixed high-level criteria to classify it. There are many ways to achieve such insight, either by examining the source of intelligence in high forms - neuroscience on human brains, or by analysing them in a representation form - as modellings, and anecdotal analogue that can be found, and so on.

As of date, no such consensus has been found about the definition of intelligence. As we have said earlier, philosopher refrains from talking of such topic, artificial intelligence practitioners rely on certain intuitive sense and reason to interpret such intelligence definition, and some argue about such notion with terms from the discipline of AI itself. In general, it is led to believe that the overall strategy is to pick on such intuition and work with it, rather than doing much about it. And it aligns well with the philosophy of defining the notion of intelligence.

1.3 Artificial Intelligence

Let us come up with an understanding of the term artificial intelligence. Make due of Stanford Encyclopedia of Philosophy (2018), it is the field devoted to building artificial animals (or at least artificial creatures that - in subtle contexts - appear to be animals) and, for many, artificial persons (or at least artificial creatures that - in suitable contexts - appear to be persons). However, such definition is fairly limited, and would not capture the essence of what practically can be artificial intelligence. Though, uncovering the current measure of which we make up artificial intelligence, we can come up with a provision definition that fits the current philosophical choice.

For the definition of artificial intelligence, or the construct that supports it, to make sense, we need to evaluate again, from what we have seen, what is even the term. As noted by definition on the notion of artificial in the preceding section, being artificial mostly comes of from the consideration of evolutionary processes - of which the interaction in the physical worlds, the biological worlds, and overall, anecdotally, of anything that is non-human of its (human) own capability to morph objects into an intended state - this is what normally resided to. Then, artificial intelligence refers to a set of observations, observable qualities deemed sufficiently of all intents and purposes intelligent, by any given constructs that is created artificially so.

This breaks down to the two conceptual ways to talk about artificial intelligence.

Conjecture 1.1 (Artificial intelligence)**.**

Artificial intelligence is the classification for any such object of constructs sufficiently reflects those qualities that fit the standard of intelligence, of which also created artificially of intent and purposes (as reflected in definition of artificial).

The second conjecture then interprets that, toward the theory of computational intelligence. If such theory is indeed proved to be feasible, then we might have the following core argument:

Conjecture 1.2.****

Artificial intelligence refers to (a)construct(s) - of which consists of the machine and its process, for such that the machine supports the process to reflects the observed results quantified in one way or another, to be interpreted as intelligence by the construct that is standard for those terms. Those constructs however, are absent, or not, by choice, of the existential facility - or of either a rigid static facility of such - and hence artificially made.

Arguably, the second conjecture is far more interesting and familiar than the first one. However, the claims, of such, can be hypothesized as perhaps not so ideal generalization. The term artificial intelligence, generically, refers to the comparison between two actual constructs. If the current human - or us - are the ones evaluating certain constructs as intelligent, then it is equivalent to generalize human into a construct on its own, of sufficient analysis such that the comparison can be conducted. If so, then the definition ultimately is reinforced, as for now, to be relative and subjective. Would we be able to find generality in such structures, if of current time we rely solely on our own construct to evaluate the criteria, though the creation that we are making is inherently different?

1.4 The Language Models (LM)

One of the main, major example of artificial intelligence application is indeed the formulation of language, manifested in a model. Attempts has been made to trying to understand why language emerges as a proxy of information exchange, either by newer treatments, as seen of Galke et al. (2022); Worden (2025), or per historical developments, as Grimm (1819); Humboldt (1836); Schleicher (1874); Darwin (1871); Saussure (1916); Bloomfield (1933); Hockett (1960); Chomsky (1957); Chomsky (1965); Goldberg (1995); Fillmore (1988); Fillmore (1988); Hauser et al. (2002); Pinker and Bloom (1990); Deacon (1997); Bickerton (1990); Kirby (2001); Christiansen and Kirby (2003); Nowak et al. (2001); Evans and Levinson (2009); Christiansen and Chater (2008); Hauser et al. (2002) and more, that is both specific of the field of linguistic and broader. The ability of forming language is considered one of the many things, up to the capacity that human is capable of, hints of intelligence that human exhibits. There, it is just natural that one of the application since the early onset of AI theory, is to recreate this form of language. Some of the first major applications are, of the onset of the Cold War, the task of machine translation (MT). The first demonstration of MT, the Georgetown-IBM experiment, showed a great promise, with limited ability that was proposed to be increased even further in the future. Though, such development did not end well, and by the time the ALPAC report came out (et al. (2006)) the field of MT has already been hit hard. It, and with the addition of Lighthill (1973) report on AI, ultimately, then officially begun the first AI winter.

Looking back as some of the failures in the theory of natural language modelling, it is perhaps surprising when looking at advancements of Natural Language Processing (NLP) has as the successor of such research direction in the prelude of AI research. Using analytical view upon the language, pragmatic approach to ‘dictionaririze’ the copula of words and sentence structures (word encoding, tokenization, data analysis-like methods), simplification of words meaning, cases, categorization, probabilistic methods (for example, Latent Dirichlet Allocation - LDA, see Jelodar et al. (2018)), such research direction is responsible toward a huge chunk of architectures, creations of ‘AI language models’ capable of statistically generating coherent texts and language contents, answering question in a sense, and so on, from large availability of data in text form. This is all conducted, while pay no mind into the deep theory of linguistic or the study of language itself; in a sense, a marvelous innovation, perhaps too marvelous. As because of such, some take the basis of the language model for the basis of the consciousness, intelligence emergence concept, and posit that such models, the LM or L (Large) LM, would be the centre piece of a fully realized AI, and thus, the discussion of AGI and furthermore, ASI. This is reinforced by the series of architectures that enables large-scale advancements, like Rumelhart et al. (1986); Schuster and Paliwal (1997); Jordan (1997); Elman (1990); Lipton et al. (2015); Graves (2012) Recurrent Neural Network (RNN), Hochreiter and Schmidhuber (1997); Gers et al. (2000); Cheng et al. (2016) Long Short-term Memory (LSTM), Cho et al. (2014); Chung et al. (2014) Gated Recurrent Unit (GRU), sequence-to-sequence model as seen in Sutskever et al. (2014), and the most foundational advanced structure of the attention-mechanism neural network — Transformer (Bahdanau et al. (2014); Luong et al. (2015); Vaswani et al. (2017)). Indeed, replicating the behaviour or coherent patterns of human language is a marvelous feat that cannot be understated. Yet, would such claim proved to be too costly, just as we have seen of criticism and empirical evidences that it is not at all omnipotent as it is pushed for?

The main fallacy of such new approach, as will be reiterated many times, is the lack of origin, and the circumstances of Descartes’s argument itself. While created such good models, it still cannot cope with logic, a wide range of logic, not simplified logic or rigid, manually designed system of logic. We are unable to determine the capacity of it to understand meaning, or any hint of such concept to exist in a language model aside from some short-lived prospect yet of no proven links to such understanding but statistical grouping. ‘Knowing language’ does not equivalent to being intelligent, as it is always said. Furthermore, there exists a transition between language of human form, words as they are being written, to numerical encoding and manual rules on such encoding law into numerical sense, that is a problem. It brings up the question of if such models, if only the algorithm that it is, does not even understand the language itself, but is just finding the best possible answer toward the task provided. Such dilemma will have to resolve, if one is to claim language model to be the standard basis of such AI generalization. As of now, that seems to not be the case, as the cracks are closer to being revealed, and as the potential AI bubble of speculation to burst and fail in such delivery. Unquestionably, again, such development cannot be understated, and should not be forgotten or relinquished. However, pushing far beyond its weight is not a good idea of such sense either.