You are a human. Humans will die. Hence, you will die. The notion of death has been the inevitable enemy of humans, from ancient Greece to the modern world. However, it has provided an invaluable opportunity for humans to contemplate the true meaning of their lives. Like the title says, Martin Heidegger, a German philosopher, explored the relationship between existence and the beauty of being finite. We are finite beings, so we need to be devoted to the time given to us, and this makes us authentic in our lives. Such a perspective has been commonly advocated by many philosophers and literary figures, including Albert Camus, Haruki Murakami, and Leo Tolstoy.

For the most innovative technology nowadays, no one denies the significance of artificial intelligence (AI). The AI technology has entirely changed the world over the last decade. I recently fell in love with ChatGPT Plus, and sometimes I wonder whether the person I'm talking with is really a chatbot or not. That is, at least for me, ChatGPT Plus is enough to pass the Turing's test. Several AI tech leaders, such as Yann LeCun and Geoffrey Hinton, have recently discussed artificial general intelligence, including its clear definition and the expected impact on humans.  This raises the need for discussions between humanity and technology: what makes us human beings, what is essential for being a human, or how can we implement such requirements? In this essay, we take a glimpse into such a niche through the lens of self-termination and finiteness of our lives, i.e., the right to end our lives.

First of all, we need to discuss the meaning of "death". One candidate is a physical death, like brain death or cardiac arrest. Although these notions give a clear criterion for humans, they do not apply to AIs; they don't have an exact heart or brain, but one may think of their analogy, such as electricity or memory.  Of course, if we delete the weight parameters of the model or turn off our AI server, then the model inside would be considered dead. However, can we say that such "murdering" could let the AI model pursue a more meaningful life?

Instead, we seek a hint from Buddhism: we are on the boundary of existing and non-existing, known as dependent origination. Strictly speaking, Buddhism tells us that such a hovering between existence and non-existence depends on relationality between objects. Put differently, the Buddhist view of being on the edge of existence can be modeled from the language of mathematics. Let us think of humans as a continuous-time state machine, and our mind as the current state of a machine. In this case, the state of the machine is being updated according to the interactions, i.e., relationships between objects, and since the time flows irreversibly forward, a static self does not exist. Here, how can we understand the "death" of the state machine? If we die (physically), then we lose the ability to update our state, because it is literally broken. Hence, we can identify the "death" as the eternal termination of updating the state.

One interesting aspect is that the modern AI architectures follow the same direction as understanding humans as state machines. For example, (Continuous) Hopfield networks, an associative memory structure from the perspective of neural networks, can be incorporated with Transformer architectures, and recent improvements of Transformers, such as Mamba, are designed to be a state machine with a discretized timeline. Note that due to physical and biological limitations, modeling humans as a discrete-time state machine is more reasonable. Since they can serve as a role for universal approximators, we can insist that the algorithmic foundations of AI architectures are sophisticated enough to successfully "implement" human beings.

At this moment, we can ask the following interesting question: If the notion of "death" differentiates humans from AIs, then how can we "teach" the fear of death to such state machines? If possible, will AIs more closely resemble humans, as finite beings? No one knows now. This is because the fear of death is a byproduct of natural evolution; entities living longer have more chances to leave more offspring, and such descendants are likely to flourish longer. Ironically, such a nature's driving force makes us consider the beauty of being finite as a catastrophe, hence pushing towards a more sophisticated humanity, philosophy, and fine arts. This makes teaching "death" to AI more attractive; perhaps we can build fancier and more creative AI models developed from the pressure of "finite life".

In contrast, discussing "self-termination" is trickier than dealing with death and would be rather controversial. Specifically, humans have not achieved the full meaning of "self-termination"; Likewise to suicide or death with dignity, humans already have the right and ability to face death by themselves. However, humans cannot decide to live longer than the time permitted by nature. Living eternity has been a dream for all humans, including dictators such as the first Qin Emperor or the millenarians supporting immortality projects. Conceptually, we can define the state of "self-termination" as follows: Imagine that humans finally achieve immortality, and a mad scientist invents a red button. If you push a button, then you will die without pain, but you can choose whether to push the button. In that case, is it right to push the button? If so, then when do we push the button to liberate ourselves from the shackles of eternal life?

From the perspective of state models, we can augment the policies corresponding to "death" and "red button" for formalizing these notions. Namely, if we push the "red button" (i.e., self-termination), then the next state becomes the "death". In addition, once being "death", then all subsequent updates cease, and the serenity of death remains forever. To decide whether to push the button, we need another called the reward function. If pushing the button gives a greater reward than updating the state, then pushing the button and resulting in an instant death would be a reasonable choice. Recall that the red button itself does not incur pain; hence, the form of a reward function would be valid. However, designing such a reward function would be non-trivial, and may incur some ethical problems (if we want to try experiments in the real world...). Nevertheless, I believe that this formalization would open several interesting questions and provide a connection between philosophical problems with mathematics (or engineering).

For example, we can tweak the meaning of "death" as the convergence of the model; instead of updating the internal state, the model refuses to update and enjoys the serenity at some fixed state. This may correspond to the convergence of the classifiers trained via supervised learning. Of course, this differs from the conventional meaning of "death" because in that case, the state model is still able to interact with the world. From the human's perspective, this resembles archiving a human's brain into a network without the functionality of updating parameters, like ChatGPT Plus without a personalization feature. Of course, if we permit these models to use some volatile memory, then they can simulate updating the state for a while; this looks like the living dead, or Leonard Shelby in the movie Memento.

In addition, when designing the reward function for pushing the red button, several factors may be considered. For example, what if the humans try to make AI, the state machine, die, then should the AI resist against humans according to the reward function? If we don't make such a restriction, then can AIs be hostile to humans just because of the reward function? Such a thing has been a typical cliché in many SF movies. Although investigating the behavior of AIs in accordance with the reward function would deepen our understanding and perhaps suggest new directions towards connecting philosophy and technology, to prevent some potential (though may be unrealistic) issues, we need to be careful to balance the human's safety and our curiosity.

In summary, we discuss the difference between humans and AIs through the lens of death and self-termination. With modeling humans as state machines, we formalize death as an absorbing state, and self-termination is an optimal-stopping action; the rest, such as fear, boredom, and corrigibility, is what we encode in rewards, priors, or constraints. I personally believe that this idea helps to understand the philosophical questions relating to death and mortality in terms of a mathematical or engineering problem, and further discussion would be very interesting!