An LLM is a massive pattern recognizer that understands and generates text.

WHYIt learns language patterns from billions of texts to generate new content.

NOWBooks, web pages, code - vast data compressed into one model.

The core of LLM is guessing 'what word comes next?'

WHYIt learned language patterns through billions of fill-in-the-blank games.

NOWHover over candidates to see the probability distribution.

Raw LLMs can't hold conversations. They become ChatGPT through Pattern Learning → Instruction Learning → Human Feedback.

WHYTransforming a pattern-only model into one that follows instructions and converses with humans.

NOWClick each stage to see how the model evolves.

Given input, it completes the answer by adding tokens one at a time.

WHYIt doesn't generate the whole answer at once, but repeatedly selects the next word.

NOWPress the button to see tokens generated one by one.

For things outside training data, it 'makes things up convincingly' - called Hallucination.

WHYThe model only learned patterns; it doesn't truly 'know' facts.

NOWInside the circle is learned knowledge; outside is unknown - it may fabricate answers.

Search 'finds' existing information, AI 'creates' new answers.

WHYThey are completely different tools. Choose based on your purpose.

NOWCompare how results differ when you ask the same question.