The NewsRupt

Training happens in stages. Pre-training exposes the model to a large text corpus and adjusts billions of parameters so the model assigns high probability to text that actually occurred. Post-training — supervised fine-tuning and preference optimisation — shapes the raw predictor into something that follows instructions and refuses certain requests. At inference the model converts your prompt into tokens, runs them through the network, and samples the next token repeatedly until it stops. Sampling settings such as temperature determine how deterministic that process is.

Ask a model to draft a refund email. It has never stored your refund email. It produces one token at a time, each conditioned on the prompt and everything generated so far, and the result reads like the refund emails that appeared in training.

Almost every practical property teams care about — cost, latency, hallucination rate, context limits — follows from this architecture. Buyers who understand that generation is prediction, not retrieval, design differently: they add retrieval where facts matter, evaluations where correctness matters, and human review where the cost of an error is high.

• That a model "knows" facts the way a database does. It holds compressed statistical structure, not records.
• That bigger is always better. Task-specific smaller models often beat frontier models on cost per successful task.
• That a model learns from your conversations by default. Most production APIs do not train on customer data unless you opt in.

• Public documentation of frontier model training data is partial, so claims about what a model has seen are usually inferences.
• Behaviour varies by model version; a statement true of one release may not hold for the next.

• Attention Is All You Need (Vaswani et al., 2017)
• Language Models are Few-Shot Learners (Brown et al., 2020)

• What is inference?