By TobiraAI – Studying Education × Tradition × Generative AI
October 21, 2025, 21:00

Greetings from TobiraAI, a humble learner of education and AI living in this corner of the world.
Thank you for reading — please sit back and enjoy this quiet study session.

🎯 Today’s goal:
To be able to explain “What is ChatGPT?” simply and clearly.

1. LLMs Are “Masters at Predicting the Next Word”

Have you ever heard the term LLM?
It stands for Large Language Model — in Japanese, 大規模言語モデル.
It might sound technical, but its essence is quite straightforward.

An LLM is trained on an enormous amount of text and learns to predict what word is most likely to come next in a given context.
Think of it as a kind of mathematical “intuition.”

That’s why it excels at summarizing, paraphrasing, and structuring text.
However, unlike an encyclopedia, it doesn’t “know” facts — it simply generates plausible sentences.
It can sometimes sound confident yet be wrong.
Understanding this boundary helps you know when to rely on AI and when to verify results.

---

2. The Mechanism in One Picture: The Transformer, a “Wide-Angle Lens”

At the core of every LLM lies the Transformer — the model architecture that revolutionized natural language processing in 2017.
It’s like a wide-angle lens, able to view all parts of a sentence at once, maintaining consistency between far-apart words such as subjects and verbs.

Earlier models had a narrow field of view, often losing coherence in long sentences.
The Transformer enabled smooth, natural writing — and it’s the backbone of today’s ChatGPT.
In short: LLMs are specialists in predicting the next words based on wide context.

---

3. What Is ChatGPT?

“GPT” stands for Generative Pre-trained Transformer.

• Generative → It creates sentences
• Pre-trained → It’s trained beforehand on huge text data
• Transformer → The wide-angle model structure

ChatGPT made this model conversational — available to everyone since 2022.
It’s not just a chatbot; it’s the result of years of progress in language understanding.

---

4. RLHF: The Etiquette Tutor That Teaches Common Sense

Even a skilled Transformer needs manners.
That’s where RLHF (Reinforcement Learning from Human Feedback) comes in.
Human reviewers rate responses as “good” or “bad,” teaching the AI to avoid dangerous or rude answers.

Through this process, a raw model becomes a socially appropriate conversational partner.
Think of RLHF as a home tutor for manners — teaching the model how to behave in society.

---

5. RAG: Bringing Evidence from Outside the Model

An LLM does not have memory like a database.
To provide accurate, up-to-date answers, we use RAG (Retrieval-Augmented Generation) — “looking up references before speaking.”

If used in government, RAG retrieves past reports and laws;
in healthcare, medical guidelines;
in factories, manuals or past maintenance records.
RAG adds traceability to AI’s words, turning intuition into evidence-based output.

---

👉 Continue reading Part 2 for practical field examples and key principles for safe AI use.