**Misconception ①: AI queries a database like a search engine**. Reality: once training is complete, an LLM (Large Language Model) encodes all “knowledge” as numerical weights in billions of parameters; during inference it queries no external databases. It’s a “compressed world model” that generates text through patterns stored in parameters. This explains why LLMs have knowledge cutoff dates (they don’t know about events after training) and why they “hallucinate” (generating information inferred from parameter patterns but not accurate).
**Misconception ②: AI is “understanding” and “thinking”**. The LLM’s core operation is **next-token prediction**: given all preceding text, predict the most likely next word. This process is implemented through the Self-Attention mechanism in the Transformer architecture, allowing the model to establish associations between words across long texts. This mechanism produces surprisingly emergent capabilities at large scale but is fundamentally complex statistical pattern matching, not “understanding” or “reasoning” in the human sense.
## Training Process Overview
LLM training has three main phases: ① **Pre-training**: self-supervised learning on massive text data (internet text, books, code, academic papers), with the goal of predicting masked words so the model learns language structure and world knowledge; ② **Supervised Fine-Tuning (SFT)**: supervised training on human-annotated high-quality Q&A pairs, teaching the model to follow instructions and generate useful responses; ③ **RLHF (Reinforcement Learning from Human Feedback)**: human evaluators score model outputs; reinforcement learning further optimizes the model to produce outputs more aligned with human preferences (more helpful, accurate, and safe).
## Why LLMs “Hallucinate”
Hallucination is one of the most important LLM limitations: the model generates information that sounds plausible but is actually incorrect. Reason: LLM’s goal is generating “statistically plausible next words,” not “factually accurate information” — when the model is uncertain, it generates “seemingly appropriate” text rather than admitting it doesn’t know (training data rarely contains expressions like “I don’t know”). Strategies to reduce hallucination: require citations (but AI can generate fake citations too); independently verify in high-risk scenarios; use RAG (Retrieval-Augmented Generation) technology combining LLMs with verifiable external knowledge bases.
See [Prompt Engineering in Practice](https://sunqi.org/prompt-engineering-guide-en/) and [3Blue1Brown Neural Network Video Series](https://www.youtube.com/c/3blue1brown).
