Office work
Reads what piled up, writes what has to be written.
Thinking that admits where it stops.
The models think before they answer: the steps happen, and what you get is the conclusion with the reasoning that carries it. Where a step does not carry, the answer says so instead of moving past it.
Three steps.
- 1
Ask the whole question
Including the part you are unsure about. A question with two readings gets a question back, not a guess.
- 2
Watch the steps
The conclusion comes with the reasoning that carries it, not after it.
- 3
Look at what it marked
Where a step does not carry, it says so. That mark is the product.
The steps, drawn out.
Decisions and processes as flow diagrams, options side by side with what speaks for each.
What it does.
- Works a problem through step by step, rather than producing a confident conclusion in one jump
- Asks back when a question has two readings, instead of picking one silently
- Separates what follows from what was given from what it recalls from training
- Says plainly when it is unsure - the layer that measures that is the reason this company exists
- Weighs options against each other, with the case against the one it recommends
- Sets maths as proper formulas and works it step by step, with a graph where one helps
Where it stops
Reasoning is not knowledge. A model that reasons well from wrong premises reaches a wrong answer beautifully. For anything current or specific, switch on web search or attach the source.
Asked like this
We have 40 000 users, 3 % churn a month and pay 1.20 CHF per user in infrastructure. At what price does the plan break even, and which assumption is the shakiest?
Which size to use
Large or Max. This is the one job where the small model is not a saving: the chain breaks in the middle and the answer sounds just as confident.
Compare the sizesQuestions.
Is this chain-of-thought?
The models reason before answering. What you get is the conclusion and the reasoning that supports it - not a transcript of every branch it considered.
Why does it say “I am not sure” so often?
Because most systems say it too rarely. A confident wrong answer costs the reader more than an honest gap, and the whole design is built around measuring which one it is.
Can it do maths?
It works through maths with the steps set as proper formulas, and a graph where it helps. Working through is still reasoning, not calculating: for figures in a file you attach it runs Python in your browser and computes them, and for anything else that has to add up, check the total.
Other kinds of work.
School and studying
Explanations that stop where the knowledge stops.
Finance
Reads the numbers, and refuses to be your adviser.