LANDMARK · 8 MIN · UNDERSTAND
Fine-tuning and alignment
After this landmark, you can distinguish training-time changes from prompt-time changes, and judge when fine-tuning is the right answer and when it is an expensive detour.
↗
This concept is shared, but the Everyday lens is less central here.
You can still explore it. We’re showing the shared explanation and a related practical view without hiding the knowledge.
After pretraining, a model is a text continuer: give it a question and it may produce a list of similar questions, because that is a plausible continuation. Turning it into an assistant takes further training. Instruction tuning teaches it to treat input as a task to perform. Preference-based training — most commonly reinforcement learning from human feedback — takes human judgments about which of two responses is better and optimises towards the preferred behaviour, which is where helpfulness, refusal behaviour, tone, and format largely come from. This stage is why two products built on similar base models feel so different. Alongside all of that sits the system prompt: instructions supplied at request time. That distinction is the one to hold, because it is constantly blurred — a system prompt is text in the window that can be argued with, crowded out, or overridden by an injected instruction, while trained behaviour is baked into the weights and applies before anyone types anything.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
The personality and the boundaries of an assistant are choices someone made, not properties of AI. A different product with a similar underlying model can refuse different things and sound completely different, and neither is the “real” one.
MAKE A DECISION
A team wants the assistant to know their 400-page internal product manual, and proposes fine-tuning a model on it. What is the better first approach?
CARRY THISAsk the same slightly sensitive question of two different assistants. Note where they differ — those differences are post-training choices, not facts about AI.