When the term Artificial Intelligence was coined at the Dartmouth Workshop in 1956, the organizers defined AI beyond narrow automata theory or simply solving problems more generally; the explicit intention was to re-create human cognition (McCarthy, 1). Unfortunately, we have considerably strayed from those goals. Today's thinking machines operate well under narrow contexts but fail to accomplish elementary human tasks. Moreover, even in those areas where they work well, they require tremendous resources, and the pace of progress is grinding to a halt (Dellinger, 2). Something has gone wrong.
So how do we reorient ourselves back to the original definition of artificial intelligence? To replicate human action, we must first understand the underlying human experience. One might be tempted to resort to neuroscience or psychology for answers, but those fields are a red herring. Heidegger instead gives us a conceptual framework to understand this human experience. If you gave the best engineer of the 19th century the parts to a Ford Model T, he would fail spectacularly at building a car. If you additionally told him what it means to drive a car, he might have a fighting chance. Neuroscience and psychology merely provide the parts; without understanding first what it means to be human, we are lost in our efforts (Jonas).
Ultimately, what we want is intelligence that can navigate, act, and interact in a peopled world: an entity that takes care of our elderly, drives our cars, builds our buildings, and keeps us company. Only entities that experience the world as we do can perform these acts. The field desperately needs a renewed Heidegerrian analysis to reorient us towards achieving artificial general intelligence.
The basis of modern AI is the neural network. A collection of neurons (layer) is stacked such that every neuron connects to neurons in the next layer. Inputs go in as numbers on one end, and some useful output comes out on the other. When numbers go from one layer to the next, these numbers are multiplied by "weights," which are assigned to connections between neurons. The values of these weights primarily determine the outcome of the network and its performance (Rosenblatt, 390). We can determine the optimal combination of these weights by performing backpropagation (see Rumelhart).
If a whole class of algorithms fails at the most basic human tasks, perhaps we should reexamine their underlying assumptions. Descartes began with the idea of an "ego" because he assumed that it is the only thing that can be known for sure ("cogito ergo sum"). The establishment of this subject necessitates the existence of an external world from which it can learn (Çüçen 1). With the subject and object established, epistemology is reformulated as such: the inner subject knows about a particular object through sense experience, then it attempts to find out about this object by observing and drawing conclusions from what is out there (Descartes, Meditation II).
The Cartesian project placed theoretical knowledge above the practical. Because the object and the subject are detached, knowledge itself becomes detached from the objective world. Let us take an example of a hammer. Under the traditional view, the mind becomes aware of itself; then, through our senses and experiences, we become aware of the hammer. When one recognizes the existence of a hammer and comes to possess theoretical knowledge about it, they can perform actions with that knowledge, thereby creating practical knowledge. One can build theoretical knowledge through deduction, observation, or perhaps an internal model about how the world operates. The critical point of the traditional view is that as a result of the subject's isolation from the world, in order to derive practical knowledge, the subject must first derive theoretical knowledge through internal means (Heidegger, 86).
Modern AI presupposes the correctness of Descartes. The model is fed a set of observations about the world from which it tries to derive theoretical knowledge through backpropagation. The neural network essentially peers out from its subjective sphere into the world and attempts to learn about it. The distinction between the subject and the object is painfully apparent. What then makes up the subjective sphere of a neural network? As a byproduct of backpropagation, the network gains a certain understanding of the world embedded in the weights mentioned above. These weights produce a "latent space" of the world which includes essential information about its observations in lower-dimensional space. For example, an image 480 pixels wide, 480 pixels high with three colors (700,000 dimensions), would be accurately represented by just 1000 dimensions when fed through a network. (Simonyan, 3).
We can draw one-to-one parallels from the Cartesian model to our neural network. Just as man is made aware of a hammer through sense-experience, the model becomes aware of objects because we intentionally feed it images of certain objects. While these networks can learn remarkably well what clumps of pixels look like which objects, if they see a hammer, it will simply show up as a collection of pixels, not as a hammer in any meaningful way.
When Ren and Lin performed a comprehensive test of a similarly structured model's ability to generate logically coherent sentences, they found that these networks perform surprisingly poorly compared to humans (Lin, 8). For example, when the network was tasked with generating a sentence when given a set of words like "dog, frisbee, throw, catch," it returned grammatically correct but logically incoherent sentences like "Two dogs are throwing frisbees at each other" (Dawson). While the model may understand what a "dog" looks like, it certainly cannot understand what a "dog" is in any meaningful way.
The AI community has previously grappled with this issue. In an attempt to imbue meaning to these clumps of pixels or sets of words, MIT researchers tried manually enumerating the use-cases for each object in the 60s (Steed, 2). More recently, networks have been stacked on top of each other in order for the second network to imbue meaning to the first network's outputs. Unsurprisingly to any Heideggerian, the Cartesian AI enterprise has been utterly unsuccessful.
If we want to make general intelligence that lives in the same world as we do, we better start with what it means to be human and what it means to live in a human world. Whereas the thinking subject was the origin of Descartes' inquiry, Heidegger begins with Dasein. Dasein, literally translated as "being there" in German, is a way of existing in the world: a mode of existence that (so far) only humans are known to experience. Dasein is characterized by our being-in-the-world, being-with, and The They.
Let us reexamine the hammer from a Heideggerian perspective. When we see a hammer, we do not see an abstract object of lines and colors. The reason that hammers show up to us as hammers is because Dasein is always concernfully engaged with the world. Dasein is involuntarily thrown into the world with its emergent projects and is forced to cope with its environment. Knowledge comes not as theoretical achievements but rather as a result of practical engagement (Heidegger, 95):
"always having to do something, producing something, attending something, making use of something, undertaking… All these ways of Being-in have concerns as their kind of being" (Heidegger, 83).
Hammers show up as meaningful to us in our attempt to accomplish these goals. If we were inert beings with no engagement with the world, what good would a hammer ever be?
Hammers also only make sense in a specific equipmental context. What use are hammers when there are no nails to nail or houses to build? We use the hammer in order to nail a nail in order to build a house for ourselves. A hammer is simply incomprehensible in a world without these things. A hammer only exists as a hammer when the "totality of equipment" exists (Heidegger, 97).
Nevertheless, we hold some theoretical knowledge about hammers; we do not simply use hammers out of pure habit. Opposite to the Cartesian model, Heidegger argues that theoretical knowledge, the spatiality or the physics of the hammer, for instance, is entirely abstracted from the ready-to-hand of practical knowledge. As a result, we usually only ever view hammers as mere objects when we encounter cases of breakdown (conspicuousness, obtrusiveness, obstinacy). If the head of the hammer came off, or there are no more nails, or the hammer is locked inside the toolbox, only then will you see it as a mere object or present-at-hand (Heidegger, 102-103). Theoretical knowledge from backpropagation should therefore not be the primary mode of acquiring knowledge.
Herein lies the problem for modern AI. Networks exist in isolation and are spoonfed observations about the world. For example, the YOLOv4 network can detect objects in an image with remarkable accuracy and speed but only insofar as it is concerned with present-at-hand knowledge (Bochkovskiy). Hence we see why our attempts to imbue meaning into the outputs of these programs through manual enumeration of use-cases or network stacking were hopeless. Genuine Dasein works the other way around, deriving present-at-hand knowledge from the ready-to-hand.
So how can we possibly simulate Dasein in artificial models? Researchers from OpenAI created a physics simulation in which networks were placed and given explicit goals such as "kick the ball into the soccer net" and "prevent the ball from going into the net" (Bansal, 4). After millions of simulations, the agents learned subtle behaviors like tacking, ducking, faking, and kicking the ball. Training techniques in which agents are placed in a virtual environment and given a task are referred to as Reinforcement Learning (RL). RL is essentially a simulation of proto-Dasein: an entity involuntarily thrown into a world with goals for itself is forced to cope with objects in the world skillfully: in this case, a ball. For all intents and purposes, the ball is meaningful to these agents. That ball is something to be shot into the goal or blocked from going into the goal. These agents do not possess theoretical present-at-hand knowledge about the ball but rather the practical ready-to-hand knowledge which allows them to use the ball in order to score goals. The resurgence and success of RL is a glorious vindication of Heidegger's formulation of Dasein (Silver). However, RL has limitations also. Applying learned models to the real world has primarily failed in sophisticated tasks.
While agents learning soccer by themselves is undoubtedly interesting, our world is more than simply two people with one ball and a soccer goal (if only things were that easy). Our world is filled with complex objects with complex meaning that require something more than just concernful engagement. Even in the simple example, we implicitly insert information about the ball and the goal into the agent's utility function from our own Dasein. Even though the agents could physically hug the ball instead of putting it through the goal, that is simply not what one does with a soccer ball. It comes to us so naturally that one plays soccer with a soccer ball or hammers with a hammer or eats with a fork, but what could be the source of this calling?
Heidegger responds with das Man (The They): a normative principle that stems from the average Dasein (Heidegger, 164). Das Man does not, however, simply correspond to the statistical average of everyday actions. Heidegger is instead describing our tendency to conform to expected behavior (Dreyfus, 153). If there is no normative principle to which we conform, there can be no equipmental context that creates meaning for objects:
"If some ate with forks, others with chopsticks and still others used their right hands… etc. whole equipmental nexus involved in cooking and eating a meal could not exist." (Dreyfus, 154).
The categories and distinctions established by das Man can only be experienced through concernful engagement with the world. For example, it is a lost cause to try to derive the distinctions of objects through science or philosophy (theoretical knowledge) simply because das Man defines these categories. It would be like trying to build a log cabin in a universe where trees do not exist. For any hammer to show up as a hammer, how one hammers, when one hammers, and what one hammers must be predetermined.
This predetermination shows up to us in the form of different characteristics of ready-to-hand objects. We can access, for example, that a dining chair is meant to be used with a table (equipmental context), that it is made out of a certain material (whereof), that it is meant for adults with large bodies (for-whom), for the purposes of eating (towards-which), which we achieve through sitting on it (use practices) (Heidegger, 97-100). These are reflections of das Man that only certain embodied beings can access. How can the use practices of a chair imbued by das Man be intelligible to four-legged beings? Of course, victims of minor injuries and disabilities remain Dasein insofar as the essential characteristics of Dasein (such as being-in-the-world) are physically possible. However, what happens as one becomes more and more detached from the physical form of the human body?
As we gradually hypothesize a body farther away from the human form, the characteristics of the ready-to-hand become more unavailable. Ultimately when you strip away all embodiment, we are left with an entity that no longer finds objects in the world humanly meaningful because it cannot access the characteristics of the ready-to-hand and, therefore, das Man.
For instance, door knobs would no longer show up as doorknobs if you had no hands. The categories set up by das Man would seem less and less intelligible to you. Let us push this logic to its conclusion: if we perfectly cloned a brain into a computer, no object in the world would show up as ready-to-hand. The meaning provided by das Man about how one eats, goes to work, and functions in society would be utterly meaningless. This disembodied brain would no longer be considered genuine Dasein. If only genuine Dasein could reason about the world as we do, we would find that no such system could abstractly reason like a human on its own.
Thus we arrive at an interesting conclusion that AI, no matter what constitutes its inner workings, even if it were a perfect copy of the human brain, would cease to be Dasein if it were not embodied. Without embodiment, there is no being-with, no das Man, no self-awareness, and therefore no common sense cognitive abilities in the Heideggerian sense of skillful comportment. This inevitable conclusion precludes popular conceptions of AGI like HAL-9000 from 2001: A Space Odyssey and most expert portrayals of AGI (see Hutter). Of course, we could still have ‘lesser Dasein’ which do not perceive the world as humans do, but if we truly wanted AGI that could exist natively in our world, the development of bipedal robotics, haptic sensors, and precise control mechanisms would be a prerequisite to the development of the brain itself.
Even so, embodied concernful engagement alone does not necessarily capture the whole truth about our ability to gain knowledge about objects. An embodied network in the wild or RL simulations still uses its weights to represent its world internally. Heidegger foresaw this mishap and warned against it. If the world, Daseins, and objects within it gain meaning from each other interconnectedly, how can networks ever hope to keep an accurate inner representation of the "whole equipmental nexus"? Tea kettles only make sense because tea leaves, water, stoves, cups, meals, and customs exist. Even if one of those factors goes missing, the meaning behind the kettle becomes irrecoverable. Dreyfus took note of this issue and claimed that human beings avoid this problem "because their model of the world is the world itself" (Dreyfus, 1140). Indeed, there is no better representation of the world than the world itself.
But how can we possibly have an agent that does not keep an internal representation? Heidegger gives us a hint by describing how humans operate:
"When one is wholly devoted to something and 'really' busies oneself with it, one does not do so just alongside the work itself, or alongside the tool, or alongside both of them 'together' (Heidegger, 405)
Instead, we do not notice the ready-to-handness of the tool at all. The readiness-to-hand must "withdraw" in order to be "ready-to-hand quite authentically" (Heidegger, 99). Now it seems like our AI should hold minimal representations of the world in its mind (as it inevitably must hold some internal representation). However, the agents should learn to skillfully cope while mostly referring to sensory inputs. This system would result in an agent that learns how to learn to live, not an agent that learns to live: it must learn how to extract meaning from das Man, the equipmental context, and its surroundings rather than embedding these things internally. For example, a traditional network would encode information about the world internally to use objects (this is tantamount to humans recognizing objects as ready-to-hand while using them, which Heidegger would detest). Instead, this new network would act mainly on the input data without utilizing internal representations.
Thus the path forward for AI is cleared by Heidegger. But by uncovering the path, Heidegger also closed what we thought were shortcuts. If artificial general intelligence with human capabilities is only possible when it is concernfully engaged and embodied, we are essentially required to create a copy of humans. Of course, AGI of 'lesser Dasein' could still be constructed and used for various purposes (as they are now). Nevertheless, AGI that can natively exist in our peopled world would require concernful engagement, lack of reliance on internal representations, and a human-like body to interpret the meaning granted by das Man.
On the other hand, Heidegger revealed that specific approaches are bound to fail. For example, methods that assume a solitary subject and a detached world are bound to fail at but the most meaningless computational tasks (such as classification or text generation). Therefore, the field should abandon the Cartesian assumptions that have previously failed them and push onwards, looking for ways to incorporate parts of Heideggerian phenomenology into concrete actionable architectures. To the best of our knowledge, this new architecture will heavily involve reinforcement learning as a training paradigm, minimize internal representations as much as possible, and be embodied. Only then will we take our first step on this new path.