LANDMARK · 7 MIN · UNDERSTAND
Reasoning and extended thinking
After this landmark, you can explain what a reasoning model does differently, why showing steps improves hard answers, and why visible “thinking” is not a guarantee of a correct one.
A plain model answers in one flow of text. A reasoning (or “thinking”) model is trained to spend extra effort before it commits — working through intermediate steps, weighing options, sometimes trying an approach and backing out — and only then giving its answer. Producing those steps genuinely improves accuracy on problems that have several stages: arithmetic, logic, multi-part instructions, careful analysis. The mechanism is still next-token prediction; the difference is that the model has learned to generate a longer working-out first, and that working-out gives later steps something better to build on. The crucial caveat is what the steps are: they are more predicted text, an explanation the model generated, not a transcript of some inner proof it checked. A chain of reasoning can look rigorous, be phrased with total confidence, and still arrive at the wrong answer — or reach the right answer for a reason that doesn't hold. So reasoning mode is a real tool for hard tasks, and the visible steps are useful for spotting where an answer went wrong; neither one turns the output into something you can trust without checking.
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
If a task is genuinely multi-step — a logic puzzle, a tricky calculation, untangling instructions — asking the model to work through it step by step, or using a “thinking” mode if the app offers one, tends to give a better answer. For a quick fact or a short rewrite it's just slower; save it for problems that actually have steps.
MAKE A DECISION
A colleague pastes a model's long, step-by-step working for a financial estimate and says, “It showed all its reasoning, so we can trust the number.” What's the most accurate response?
CARRY THISTake one recent task where you used AI. Decide whether it was genuinely multi-step (worth reasoning mode) or simple (not), and note one way you would check the final answer independently of the model's own explanation.