Go for technicality, neural network architecture of itself faces several increasingly difficult dilemma. Let us disregard the problem of Neural Scaling Law, Kaplan et al. (2020b), as it is simply an empirical optimization observation hyped of its wording[^5], the theoretical ground on which neural network is born from, and the empirical method it employs to advance ever since, face problems regarding interpretability, structural effectiveness or definition, phenomena explanations, uncontrollable behaviours, black-box restrictions, rigidness of architecture, uncertainty and vanishing problems, overloading problems, and so on. Interpretation is the largest problem with such theoretical and both heuristic treatment, as no one understand what lies underneath such system and the operation that births such result observed or tasked itself. Explainable AI, like presented in Hsieh et al. (2024), was created to counter such issue, still cannot work or cannot make substantial advancements aside from careful manual designs and limited domain analysis, which in an unfair bit of comparison, go back to the time of symbolic AI approach. This is even worse for neural network, as the regular wisdom is that no one understands the operation of hidden layers and all, in any given setting, aside from which you reduce it to very small size like a perceptron or so. Theoretical development toward such is also slow, and often simplify it to mathematical object to be analysed too rigorously, losing the essence of the architecture in the mathematics rigours. This is particular event or rather pattern that we unfortunately stretched to the previous section on learning theory. Often see in learning theory development, or machine learning theory in general - the ‘mathification’ of a theory is a problem that somewhat plagues papers and researches on machine learning topics. Even though we need mathematics on either end, the approach is impractical or simply wrong, believing mathematics to be the singular thing that defines learning theory. This has many fallacies that can then be attributed to a lot of factors and whatnot, of several factors that plague this analysis even further than just the problem of double descent. For now,
- Epistemological limits and interpretations (Barbierato and Gatti (2024)) - We ultimately lack understanding of a lot of things. While those ‘theorems’ are very nice in learning theory, the real picture is that it is not real learning, for the word learning are not even defined, as such is to compare them to human learning on itself. Even by then, theorems are severely limited. Certain voices also concern of similar problem, including Lipton (2018); Doshi-Velez and Kim (2017); Molnar et al. (2020) (Molnar et al. (2020) shortly consider the misleading interpretation question instead), and philosophically, with several pushbacks on structural anecdotes, Dreyfus (1965); Dreyfus (1972); Dreyfus and Dreyfus (1986); Suchman (1987); Brooks (1991); Searle (1980b); McCarthy and Hayes (1969); Harnad (1990a). On the more modern side, of contemporary critiques, Pearl (2009); Marcus (2018); Sutton (2019). Of the Chinese Room Argument, perhaps we can look into already copulated passages.
- Applying wrongly, and is used to impress and not to explain anything (Lipton and Steinhardt (2018)) - mathematics is used to impress certain demographic of reviewers and readers, to provide a sense of rigours, to further enhancing the image of formal theory to the point that such theory, even if wrong, can be considered fairly correct by the sheer volume of practitioners believing in such. Such is to say the mathiness is turning things into ideology more than rigours itself, of which we can take a tangent to see in economic theory, one of the place to adopt a large portion of machine learning statistics, the pushback against such cursed devolution (Romer (2015); Syll (2024)). Such can also be seen in generally most double descent analysis, in which there exists many formalisms yet not definite result on double descent.
- Reproducibility and acute false claims (Kapoor and Narayanan (2022)) - In general, what we have done cannot be recreated, in certain way, and of certain too optimistic setting that is seemingly unrealistic - particularly in adoption toward practical means of actions. Certain theoretical assumptions and formulations are too strict, of which means the increment of hypothesis fluctuation with removal of such constraints.
Continuing, practical problems range from grokking (Power et al. (2022); Davies et al. (2023)), double descent (Belkin et al. (2019); Nakkiran et al. (2019)), triple descent (d’ Ascoli et al. (2020)), counter-intuitive results (Szegedy et al. (2014)), adversarial examples or statistical instability (Goodfellow et al. (2014); Carlini and Wagner (2017)), ‘memorization issue’ (Arpit et al. (2017)), generalization problem and its purposes (Zhang et al. (2017)), catastrophic forgetting (Kirkpatrick et al. (2017)), and so on. Terms like inductive bias are thrown in of many contexts without a single grounded definition or notion to support such, many terms where born without actual consideration, concepts are thrown in and out without basis. While it is easy to simply ignore such problems, such as to be seen in usual deep learning practices of AI, the crack inevitably shows, and a very wide range of developmental gap ensues.
2.4 The end goal of AGI
With that said, what is the end goal of AGI? In all of it, the central goal is supposed to be the creation, out from the doctrine of fragmented intelligence, a unification in which we can then centralize everything in a singular artificial intelligence system, hence AGI. However, as we stated, the term AGI can be further dissected to subsets of such AGI construct, and even then, different system has different AGI threshold. Thereby, the end goal is not clarified yet. What we are seeing instead is a wave of hysteria, from people afraid of the incremental development, and the current market bubble generated from the present AI boom, all of which in turn, pushes the notion of ‘building AGI’ further and further to the truth. Such is also amplified of the outlook to the concept of ASI, of which is sold in the general media as the next stage on which post-scarcity can be achieved, where Nobel-level researches are outputted every few days, and automation leaves human of no burdensome tasks. However, with what is being done right now, it seems like such outlook cannot be realized, at least in due time.
The conflict in which the philosophy of which promised AGI to the public, as well as the optimistic outlook of people and thinkers, perhaps can be reflected in the sense of Heideggerian philosophy, or famously so, the philosophy between ready-to-hand (Zuhandenheit) and present-at-hand (Vorhandenheit) (Heidegger (1962); Dreyfus (1991)). Heidegger simply explains, of the notion in which the object appears in front of us. In a sense, the subject of question, and the person of inspection, of all but anthropocentric. For Heidegger, most of the time we are involved in the world in an ordinary way or ”ready-to-hand.” We are usually doing things with a view to achieving something, and hence, with a purpose already configured, and project such voice outward. The being of the ready-to-hand announces itself as a field of equipment to be put to use, and hence ultimately defines itself around such action and potential. A famous example of such notion is the hammer analogy. In question toward what the hammer can do, whether because from intuition or else, we extrapolate such object of its functionality, instead of inclining on the more sophistry of decompositions and analysing components. Such knowledge can then be made trivial, or thus transferred to a lower priority of thought, of which to the point where one can talk comfortably of other subjects or adjacent topics, but do not require careful contemplation to the original consideration (the hammer). The reversal of such action happens in isolation, in a typical understanding of which decomposition is required, structural knowledge needs to be formed, and so forth, such is to say we purposefully study such object similar to a scientist in idealization studies carefully of certain object, and hence, defers its meaning and existence to the notion of it ”exist” to be there. That is the idea of present-at-hand.
It is then fairly connective to approach from certain angle, that concerning Heidegger’s notion, our AGI system, plus the non-elucidating way of studying the architecture of choice for such endeavour of AGI, we deliberately refuse the notion of present-at-hand for a more ready-to-hand, in some cases to the extreme of empiricism, of which then also founded the principle of agentic AI system. This can be applied to even further, as to the normal principle and conceptual architecture of artificial intelligence system in an agent-environment scheme as being inadequate, correct, but not enough (Stanford Encyclopedia of Philosophy (2018)). There will be, ultimately, a wall in which this approach cannot reach, as to there exists too many permutations of a path such that there exists fundamentally, of a statistical approach, to be an almost zero chance in which such can happen, within the wrong mode of understanding, both from the perspective of the designer, and the perspective of the construct itself. In the same spirit of the UAT theorem, we can say that the optimism will be, to a certain point, similar to saying ”Somewhere in the space of all possible English sentences, one perfectly describes quantum gravity.”. Interesting and true (probably), but useless.