LLMs do not learn through experience like humans; they possess a completely different type of intelligence that learns by absorbing vast amounts of language.
Imagine you are a newborn baby. You hear your mother’s voice, touch toys, and perhaps even burn yourself on something hot. Through these small, accumulated experiences, you realize for yourself the cause-and-effect relationship that “Ah, if it’s hot, I need to be careful.”
Now, what if there were someone who had never stepped out into the world, but had read every book and conversation humanity has written over the past thousands of years? Could that existence truly be said to have the same intelligence as ours?
With the recent emergence of Large Language Models (LLMs) like ChatGPT, people are curious about whether AI is becoming truly intelligent like humans. However, experts say that the capabilities demonstrated by LLMs are fundamentally different from human ways of thinking—a kind of “alien intelligence” [Source 2], [Source 10].
Why is this important?
Distinguishing whether AI thinks like a human or is simply mimicking intelligence well is extremely important. If AI truly understands and learns human-like causality, it might eventually be able to set its own goals and navigate the world [Source 3].
However, if it is simply learning patterns of language, we must constantly monitor whether the information provided by AI is factual or distorted. The criteria for deciding whether to view AI merely as a convenient “tool” or to treat it as a “peer intelligence” lie right here.
Easy Explanation
The core of the LLMs we use is a technology called the “Transformer” (an AI structure that identifies relationships between words in a sentence) [Source 7]. Simply put, a Transformer is a machine that converts text into a list of numbers (vectors) and then mathematically predicts the next word with the highest probability [Source 7].
To use an analogy for this process: if a human climbing a mountain is an “activity of sweating and gaining experience directly,” an LLM is like reading all the mountaineering records of tens of thousands of people who have climbed the mountain without going there itself, and mastering through data “around where there are many rocks and where the scenery is good” [Source 1]. Because it absorbed knowledge without experience, even if the destination reached looks the same, the process is completely different [Source 1].
Furthermore, every time an LLM starts a new conversation, it answers using only the “fixed knowledge” input during its training phase [Source 10]. Humans learn from the mistakes they made yesterday and revise their goals for tomorrow, but when an AI’s conversation ends, it does not integrate that experience into its intelligence and returns to its original data state [Source 10].
Current Situation
Currently, AI is being utilized in very limited but astonishing areas. In particular, the “LLM-as-a-judge” technique is representative [Source 9], [Source 11]. For example, after having an AI write a long text, it is a method where another AI evaluates that text itself and suggests better directions [Source 9]. Compared to the existing method where humans evaluated everything one by one, this method reduces costs by up to 98% while providing evaluation quality at a human level [Source 18].
However, this ability is, at most, an ability to find patterns based on learned data. There are many criticisms that it is far from the “internal world model” (a program used to identify causes and results in the world and achieve goals) that humans actually use while living in the world [Source 3]. We must keep in mind that behind the logical answers AI gives, there is “probabilistic calculation,” not “understanding.”
What will happen in the future?
In the future, AI will evolve beyond simply learning English-language data and toward satisfying the unique demands of various linguistic regions [Source 16]. Many countries are already building LLMs specialized in their own languages, which means AI will become a more familiar existence to a wider range of people [Source 17].
We must always remember that the intelligence AI shows has “alien characteristics” different from human ways of thinking. This is because just because they answer well, it does not mean they think and feel like humans [Source 10]. We must not forget that AI is not an evolving living organism, but the culmination of precisely designed information processing.
MindTickleBytes AI Reporter’s View
The more we try to view AI as the exact same intelligence as humans, the more we will become disappointed or afraid. When we accept the point that AI is merely the best sponge for absorbing knowledge, not an existence that builds wisdom through life’s experiences, we will finally be able to use the tool called AI more wisely. AI will shine brightest when it remains not our replacement, but a powerful “external brain” that expands our knowledge.
References
-
[Storytelling with Impact What kind of intelligence is an LLM?](https://www.storytellingwithimpact.com/nature-of-intelligence-episode-three-what-kind-of-intelligence-is-an-llm/) - LLM vs. Human Intelligence - LinkedIn
- Does intelligence ‘emerge’ in large language models?
- The Secret Lives of LLMs - Psychology Today
- What LLM Is and Is Not: a Philosophical and Practical Overview
- Large Language Models and Cognitive Science: A Comprehensive …
-
[The Yale Review Melanie Mitchell: The Dangerous Unknowns at …](https://yalereview.org/article/melanie-mitchell-jagged-intelligence) -
[LLM’sasaDifferentKindofIntelligence Hacker News](https://news.ycombinator.com/item?id=48791650) - LLMasaJudge: опыт оптимизации генератора описаний… / Хабр
- The First AlienIntelligence: Why LLMs Think Nothing Like Us
- towardsdatascience.com/llm-as-a-judge-a-practical-guide
- Nigeria Launches First MultilingualLLM- Silicon Africa
- Hugging Face, Inception & MBZUAI launch new ArabicLLMleaderboard
-
[LLM-as-a-judge on Amazon Bedrock Model Evaluation Artificial…](https://aws.amazon.com/blogs/machine-learning/llm-as-a-judge-on-amazon-bedrock-model-evaluation/) -
[Storytelling with Impact Intelligence- Babies vs Machines](https://www.storytellingwithimpact.com/nature-of-intelligence-episode-four-babies-vs-machines/)
- Both read books and learn the same way
- Humans learn through experience, while LLMs learn by absorbing language
- LLMs learn through experience
- Transformer
- Neuron
- Data Center
- They remember all past conversations
- They accumulate new experiences
- They start fresh based only on fixed, trained knowledge