A chatbot does not look an answer up. It predicts the next word, adds that word to what it has read, and predicts again. Each prediction is ordinary arithmetic: a table lookup, a few thousand multiply-and-adds per word of input, a normalization, and one exponential for every word in its vocabulary. A large model runs the same arithmetic on vectors thousands of numbers long, which is why nobody follows it by hand. The model in the figure is small enough that you can. It reads a 107-word vocabulary, carries 16 numbers per word through 2 blocks of 2 attention heads, and has 9,984 trained numbers in all. It was trained on 185 simple questions (capitals, colors, animal sounds, small sums and opposites), and it answers all 24 wordings that were held back from its training.
Every word starts as a row from a lookup table plus a second row for its position. From there, each word's vector rides along one line through the network, the residual stream. Each stage reads the vector, computes a change, and adds the change back, so the thick line in the map is the same 16 numbers being revised four times.
Attention is the only stage where words look at each other. Each word makes a query, a key and a value. The dot product of one word's query with another word's key scores how relevant the second word is. Softmax turns a row of scores into weights that add up to 1, and the word's output is the weighted average of the values. Ask "what is two plus three ?" and open step 3.3: on the pass that writes the answer, one head in block 1 puts 99% of its attention on "two" and the other puts 99.8% on "three".
The MLP comes after attention and works on each word separately: up to 64 neurons, a GELU that passes positive values and squashes negative ones, and back down to 16 numbers. After the second block, the last word's vector is scored against all 107 words. Softmax turns those scores into probabilities, and the model takes the top one. Then it appends that word and runs the whole calculation again. "it", "is", "paris", "<end>" is four complete passes and about 250,000 multiplications.
The model only knows the patterns it was trained on. Ask "what sound does a banana make ?" and it answers "yellow" with 30% confidence, because the one thing it learned about bananas is their color. Ask "what is the opposite of france ?" and it says "paris" with 80% confidence, which is a reminder that confident and correct are different things. Words outside the vocabulary are refused; a real model splits an unfamiliar word into smaller pieces it does know. Large models use the same operations at a far larger scale, plus refinements this figure leaves out. The longer version, with the scale comparison, is the article The Arithmetic Inside a Language Model.