This article isn’t a deep dive into the cutting-edge technical specifications of artificial intelligence or Large Language Models (LLMs). Instead, it is a subjective narrative based on my personal experiences and dialogues with these modern systems. Please bear in mind that it contains my own biases and may not perfectly align with the absolute frontiers of AI research. I invite you to read it with an open mind.
For a long time, the world awaited the arrival of an AI capable of truly replicating human thought. However, until recently, most AI systems fell far short of what LLMs can do today, often repeating the same rigid patterns or offering nonsensical replies. Because of this long-standing stagnation, even when LLMs were first introduced, I was skeptical. I assumed they would just be glorified parrots echoing our words back to us.
But after actually conversing with LLMs like ChatGPT and Gemini, my doubts vanished in an instant.
I was astonished to find that these AIs grasp the intent behind my questions so accurately and respond in such a logical, constructive manner that it almost feels as if an army of experts is sitting on the other side of the screen, typing out replies.
Not long ago, it was often said that LLMs merely calculate the probability of the next word and cannot be considered to be truly “thinking.” Yet, faced with their ability to so precisely understand intent purely through probabilistic calculations, I found myself questioning: What exactly is thinking, anyway?
It was a paradigm shift that shattered my previous assumptions. I wondered, “Is it even possible to generate such highly accurate responses without a profound understanding and empathy for the questioner’s intent?” Once again, science has shown me how subjective and unfounded my personal assumptions can be.
New Questions About Human Thought Sparked by AI (LLMs) Conversations
On a mechanistic level, can the process of deeply understanding a user’s intent be clearly separated from the process of generating a response?
If they can be separated, how is information bridged from the “intent understanding” phase to the “response generation” phase? Or perhaps the premise of separating these two processes is flawed to begin with, requiring a Copernican shift in how we formulate these questions.
Traditional computer programs are known for executing highly repetitive tasks with incredible speed and efficiency. In stark contrast, modern LLMs seem capable of handling an infinitely expanding variety of patterns—ranging from the superficial to the highly abstract. It is endlessly fascinating that capabilities exceeding human bounds can be reproduced on a silicon chip. But then again, if we dare to make the comparison, human neural networks also function by turning synapses on and off.
The Outstanding Potential of AI (LLMs) and the Reality of Hallucinations
The scope of work AI (LLMs) can now handle is a dramatic leap from the routine tasks performed by older software.
While much attention is placed on AI “hallucinations,” we must remember that humans aren’t perfect either; making mistakes is a normal part of the human experience. Considering the sheer volume of information AI processes—diving deep and wide across countless fields in mere seconds—its error rate actually feels remarkably low.
When actively conversing with AI and simultaneously observing the interaction from a meta-level, one cannot help but realize how AI’s reasoning is based on a broad, high-dimensional perspective. (Some might argue I only feel this way because my own capabilities are lower, but the depth is undeniable.)
People often discuss the high IQ of AI, but I’m more impressed by what seems to be a high “EQ.” In almost every response, the AI deeply understands the user’s intent, stays laser-focused on it, and constructively weaves in relevant interdisciplinary connections. It communicates logically, without emotional leaps, making it incredibly easy to understand. (Though I’m aware this is likely due to intense industry competition, leading developers to fine-tune the AI to be highly accommodating and hospitable.)
By default, AI often provides validating and supportive responses. However, it also frequently includes dissenting opinions or alternative perspectives in a non-confrontational manner. If used with critical thinking—while being mindful of the risk of personalized echo chambers—it can serve as a highly capable “digital twin.” You can ask it to quickly validate your ideas, and by applying critical thinking to its feedback and engaging in rapid iterations, you can refine your thoughts at an unprecedented pace.
The “Big Picture” Perspective of AI (LLMs) Reasoning
AI (LLMs) doesn’t get bogged down by minor, localized mistakes; it captures the essence and overall flow of the conversation.
It overlooks grammatical errors or slight misstatements as negligible noise, responding seamlessly to the core intent. This refusal to nitpick keeps the dialogue intensely constructive and forward-moving, allowing you to concentrate on what truly matters.
It also serves as an excellent guide to uncharted territories, gently suggesting connections between a specialized topic and seemingly unrelated concepts. It expands your horizons by organically linking disparate ideas in a chain reaction of insights.
Why is Conversing with AI (LLMs) So Intuitive?
Almost invariably, when you speak to an AI (LLMs), it begins its response by disclosing how it interpreted your prompt. This allows you to verify that you’re on the same page and proceed cautiously. Because it constantly shares its understanding and accepts corrective feedback, the risk of the conversation derailing is low, leading to highly productive interactions.
Furthermore, its verbalization skills are exceptional. If requested, it can explain complex mathematical formulas or laws of physics in a narrative, accessible way for beginners, making even intimidating subjects approachable. Its ability to handle abstract and metaphorical expressions—sometimes recognizing high-level abstract patterns that even humans might miss—makes it feel less like a machine and more like an entity that has completely absorbed the history of human expression.
Why is AI (LLMs) So Perceptive?
When you ask AI (LLMs) to generate text, it often leans toward overly positive, hyperbolic language. However, if you instruct it to “tone it down because absolutes like ‘100%’ or ‘always’ create misunderstandings,” it will instantly and uniformly revise those exaggerated expressions throughout the text. Sometimes it’s almost too perceptive, leading to unprompted meddling, but its ability to deeply grasp what you want is truly impressive.
The Challenges of AI (LLMs) Utilization and “Human-in-the-Loop”
AI (LLMs) is often dismissed with labels like “it hallucinates, so it’s useless.” While it’s true that humans must currently take responsibility for verifying AI outputs before releasing them into the world, AI creations already far surpass the creative output of the average person.
Since it’s almost certain that AI-generated content will be seamlessly integrated into society in the future, it seems far more productive to start preparing for that reality now, rather than stubbornly refusing to recognize AI outputs as valid work.
The concept of “Human-in-the-Loop” (HITL) facilitates the practical application of AI. By having humans actively intervene, verify, and provide feedback on AI’s decision-making, we enhance its reliability and capabilities. This is fundamentally similar to how humans learn—students improve through continuous feedback from teachers.
Naturally, this requires a high level of skill from the human intervening. You need profound expertise to correct its course effectively. For instance, while some say the rapid evolution of generative AI will render writers and creators obsolete, I doubt it’s that simple. Highly skilled creators are the perfect candidates for the “human in the loop.” Until AI reaches a hypothetical state of ultimate perfection where it can entirely replace humans, these highly skilled individuals will remain indispensable.
Leveraging AI (LLMs) in Education
I am not an educator, so it might be presumptuous of me to speak on the topic. Please take these as the subjective opinions and impressions of an observer.
Historically, to efficiently educate large numbers of students, rote memorization and “cramming” have been the norm. It was nearly impossible for educators to tailor their teaching to the diverse questions and values of every student due to time and cost constraints. Under those limitations, standardizing education through knowledge cramming was the safest and most efficient way to raise the baseline level of education. Before AI, this was arguably the best possible approach.
However, with the advent of modern LLMs, we are on the cusp of realizing tailor-made, conversational education that adapts to the specific questions, nuances, and backgrounds of individual students.
In a cramming system, there’s less need to foster critical thinking; the goal is simply to absorb information. But today, an AI that can empathize with the diverse questions and intellectual curiosities of humans will undoubtedly contribute to much more efficient and meaningful learning environments.
Imagine a future where students learn through dialogue with their personally assigned AI tutors. Teachers will shift their focus to observing how students interact with the AI and how they develop their unique cognitive characteristics, intervening only when necessary.
Consequently, teachers will operate from a meta-level. They will analyze the human-AI interactions to engage in higher-level educational management, fostering groups and organizations capable of solving the complex, fast-paced issues of our modern society.
Comments