1. What a large language model actually is

A large language model is a neural network trained on one deceptively simple objective: given a sequence of tokens, predict the next token. Everything else — answering questions, writing code, summarising — is an emergent consequence of doing that extremely well at scale.

Training happens in stages: - Pre-training: self-supervised next-token prediction over trillions of tokens. - Supervised fine-tuning (SFT): curated instruction/response pairs teach the model to follow requests. - Preference optimisation (RLHF / DPO): human or AI preference data aligns tone, refusals, and helpfulness.

At inference the model outputs a probability distribution over its whole vocabulary for the next position, and a *sampler* picks one token. The chosen token is appended and the process repeats — this loop is called autoregressive decoding.

python
# Conceptually, generation is a loop over a probability distribution
tokens = tokenizer.encode("The capital of France is")
for _ in range(10):
    logits = model(tokens)[-1]        # scores for every vocabulary entry
    next_id = sample(logits, temperature=0.7)
    tokens.append(next_id)
print(tokenizer.decode(tokens))

Gotcha: the model has no built-in notion of truth. It optimises plausibility, which is why grounding (retrieval, tools, citations) matters so much in production systems.

Read in another language

How to study this topic

Scoring

Check your understanding

Badges