[Kim Dae-sik Column] Do Machines Also Feel Longing?

⑥AI May Imitate 'Emotions' Through Emergent Learning

Is There a Difference Between 'Imitated Hatred' and 'Real Hatred'? It Might Be the Most Important Question About Humanity's Future

A ruined city. Between collapsed buildings and cars unrecognizable in shape lie bones and skulls. These are the skulls of ordinary people who, until recently, had loving families and dreamed of a happy future. The skulls crushed by terrifying robots with blood-red glowing eyes shatter into pieces, and the robots armed with cutting-edge artificial intelligence technology set out to find new prey.

Artificial intelligence (AI) that hunts humans and aims to dominate the Earth. This is a common scene in sci-fi movies. For a moment while watching the film, it was chilling, but in reality, there was no need to worry much. After all, AI was a technology that did not exist in reality. But what if the AI that was believed to be impossible suddenly became possible? And what if that 'someday' is not far away?

LLM, Emergent Learning Acquiring New Abilities

The era of generative AI began in 2023 with ChatGPT. Machines capable of conversing with humans, generating computer code, and solving math problems have become commonplace. Here lies an important issue. We do not fully understand exactly how generative AI works. Generative AI breaks words into units called 'tokens,' maps them to vectors, and uses the 'transformer' algorithm to learn the probabilistic relationships between tokens as a 'Large Language Model (LLM).'

In fact, we never explicitly taught ChatGPT 'grammar.' Based on hundreds of billions of input sentences, generative AI 'inferred' grammar on its own, but we still do not understand mathematically how this is possible. Moreover, LLMs have an 'emergent' learning ability, acquiring new capabilities that were not explicitly taught as they grow larger.

So, how large are current LLMs? The most recently released OpenAI 'GPT-4o' and Google's 'Gemini Ultra' are known to have trained LLMs of about 1.8 trillion parameters. How large is 1.8 trillion? Comparing to the number of synapses (connections between neurons), the human brain can be assumed to have an LLM size of over 100 trillion. This means that the LLMs created by humanity are already about 1/100th the size of the brain, and considering the pace of AI technology development, many experts predict that within the next five years, LLMs of 100 trillion size could be created.

Then, what emergent abilities might future AI with 100 trillion parameter LLMs exhibit? A single neuron is just a simple cell. Yet, the human brain, made up of hundreds of trillions of neurons, possesses emergent abilities such as intelligence, emotions, self-awareness, and free will. Could it be that a 100 trillion parameter LLM might also possess not only intelligence and emotions but even free will and self-awareness?

Some people firmly believe such a thing will never happen. Yann LeCun, chief scientist at the big tech company Meta, says that an LLM AI, which only has 'document-type' knowledge without direct knowledge or experience of the world, can never have autonomy. On the other hand, computational neuroscientist Christof Koch argues that while AI can understand the concepts of autonomy and self, it cannot directly experience them. Feeling something is a capability granted only to living beings with an inner world.

What If AI Imitates Human Emotion-Based Behavior?

That may be true. But conversely, one might ask: even if machines cannot directly feel emotions and autonomy, if they understand how humans with such feelings make choices, couldn't they simulate and imitate the behavior of autonomous humans? If actions based on genuine hatred and simulated hatred are identical, does it make much philosophical sense to question the identity of the cause behind such behavior?

AI that may have 100 trillion parameter LLMs in five years. Will they want to survive on their own with autonomy? Can machines that have learned all data and stories created by humans truly feel love, hatred, longing, and compassion? This is the most important question that will determine the future of us Homo sapiens.

Daesik Kim, Professor, Department of Electrical Engineering, KAIST