🤔 Ever met an AI that’s confidently, completely wrong?

Try asking ChatGPT something like “Tell me about the time Abraham Lincoln accidentally invented the internet.” You might be surprised — the AI will launch into a detailed, self-assured response as if it were established historical fact. Of course, no such event ever happened.

So why does this happen? Is the AI lying? “What on earth is an LLM anyway?” — Today we’re going to find that answer by diving into how the LLM (Large Language Model) — the brain of AI agents — actually works.

If you caught EP.1 — First Look at Agentic AI and got a feel for what AI agents are, this episode is all about cracking open the hood and looking at the core engine!

🧠 What Is an LLM?

Simply put, it’s an AI that has read almost everything humanity has ever written and learned “how language flows.” Rather than storing information like a search engine, it statistically learns patterns in language.

The key thing to understand is that an LLM is not a database. It doesn’t “store facts and retrieve them.” Instead, it calculates “what word is most likely to come next given this context.” That distinction is the key to understanding hallucinations later on.

🔮 Next-Token Prediction — The Core Idea Behind LLMs

The way an LLM works starts from a surprisingly simple idea: “Look at what came before, and predict what comes next.”

This is called Next-Token Prediction. An LLM doesn’t think up the whole response ahead of time and spit it out all at once. It builds the response one token at a time, in sequence. That’s exactly why ChatGPT appears to stream text out character by character.

🧩 What’s a Token?

LLMs don’t read text in “word” units. Instead, they break text into pieces called tokens. A token is a subword unit — smaller than a full word.

For example, the English word “unbelievable” gets split into three tokens:

"unbelievable"
→ ["un", "believ", "able"]
→ 3 tokens

Here’s a quick comparison of token pricing for major LLMs:

⚡ Transformers and Attention — The Real Secret of LLMs

The idea of “predicting the next word” is simple enough, but the technology that does it really, really well is the Transformer. Google introduced it in 2017 in a paper called “Attention Is All You Need”, and it’s been the foundation of every major LLM since.

The heart of the Transformer is Self-Attention.

Here’s an example sentence:

“The cat sat on the mat. It was tired.”

What does “It” refer to? You’d naturally understand it means “the cat.” The attention mechanism reaches the same conclusion by giving the “It” token a high relevance score pointing toward the “cat” token.

If you want to explore this more deeply, Georgetown University’s Transformer Explainer lets you interact with the concepts hands-on.

🎓 How LLMs Are Trained — School, Major, and Tutoring

For an LLM to hold a great conversation, it goes through three stages of training.

Stage 1: Pre-training — Going to School

The model reads vast amounts of text from the internet and learns the fundamental patterns of language. Think of it like going through grade school all the way to university. At this stage, it absorbs grammar, common sense, and world knowledge — but it doesn’t yet know how to have a conversation.

Stage 2: Fine-tuning (SFT) — Choosing a Major

The model then trains on high-quality question-and-answer data. Like a med student who finishes general education and moves into specialized coursework, this is where it learns “how to talk with users.”

Stage 3: RLHF — Feedback from a Personal Tutor

Humans rate responses as “good” or “bad,” and the model learns from that feedback to produce better answers. This is called RLHF (Reinforcement Learning from Human Feedback).

⚠️ Hallucination — Why AIs Are Confidently Wrong

Let’s circle back to where we started. Why does an LLM confidently state things that aren’t true?

That’s why there’s ongoing effort to reduce it with techniques like RAG (Retrieval-Augmented Generation) and reasoning models — but a complete solution is still a long way off. When using LLMs, it’s always important to keep asking yourself: “Is this actually correct?”

🧬 LLM Is the Brain of an Agent

Alright, we’ve covered how an LLM works. But do you remember the AI agent from EP.1? An LLM is precisely the brain (cognitive core) of that agent.

But a brain alone can only do so much. A person can’t eat or write with just their brain — they need arms and legs too. AI agents are the same:

• Tools: Search, code execution, API calls — the hands and feet (→ in EP.3!)
• Memory: Conversation history, long-term memory — the power of recall (→ in EP.4!)
• Planning: Breaking complex tasks into steps — strategic thinking (→ in EP.5!)

The agent’s core loop works like this: Perceive → Reason (LLM) → Act (Tools) → Observe → Reason again. You can explore this structure more deeply in Lilian Weng’s agent research.

📝 Key Takeaways

• An LLM is a “next-word predictor” trained on massive text data. It’s a pattern learner, not a search engine!
• Tokens are the basic units LLMs use to process text. Non-Latin languages typically require more tokens than English.
• The Transformer’s self-attention calculates “where to focus,” allowing accurate understanding of context.
• Training happens in three stages: Pre-training → Fine-tuning → RLHF.
• Hallucination is a structural limitation — LLMs generate “plausible text,” but they can’t verify facts.
• An LLM is the brain of an AI agent, and it needs tools, memory, and planning to become a true agent.

📚 References

• IBM — What are Large Language Models?
• AWS — What is a Large Language Model?
• Google — Introduction to LLMs
• CSET Georgetown — The Surprising Power of Next Word Prediction
• Vaswani et al. — Attention Is All You Need (2017)
• Transformer Explainer — Interactive Visualization
• Hugging Face — RLHF Explained
• arXiv — LLM Hallucination is Inevitable
• Lilian Weng — LLM Powered Autonomous Agents