How Does ChatGPT Actually Work?

Frame the driving question by showing a familiar chat exchange and asking what could possibly be producing the reply.

• ChatGPT writes fluent, context-aware replies
• It is not searching the web or looking up a database
• Something inside the model is generating each word

Let the learner commit to one intuition before the mechanics are revealed.

• Choose the model that best matches your current intuition
• Your guess will be revisited after the explanation

A simple simulation that, given a few prompt words, shows the probability distribution over possible next words so the learner can see prediction in action.

• The model assigns a probability to many possible next words
• One token is sampled from that distribution
• The chosen token is appended and the process repeats

Show how text is broken into tokens, the actual units a language model reads and writes, with a visual of a sentence split into colored token blocks.

• Models do not see letters or words directly
• A tokenizer breaks text into common subword pieces
• Each token is converted into a vector of numbers

An interactive diagram where the learner highlights which earlier tokens a given token attends to, revealing how context flows through the model.

• Attention lets each token weigh the relevance of every other token
• Multiple attention heads run in parallel
• This is the core mechanism behind long-range coherence

Walk through pretraining on broad text and then reinforcement learning from human feedback, explaining why two stages are needed.

• Pretraining teaches general language patterns at huge scale
• Human feedback teaches helpfulness, safety, and style
• The same next-token objective drives both stages

Let the learner adjust a temperature slider on the same prompt and compare the sampled outputs, testing the idea that randomness shapes style but not knowledge.

• Low temperature yields predictable, repetitive text
• High temperature yields varied but riskier text
• The underlying knowledge distribution is the same

Address common misconceptions: ChatGPT does not store facts as a database, does not think in human sense, and can confidently produce wrong information.

• It has no live search by default
• It has no internal truth-checker
• Confidence in tone does not equal correctness

Tie everything back to the driving question with a single, clear answer and revisit the learner's earlier intuition.

• It tokenizes your prompt and predicts the next token, repeatedly
• Training on text plus human feedback shapes those predictions
• Scaled up, this simple objective produces helpful, conversational replies