aime now thinks more carefully before it answers
Our newest reasoning model holds its ground on tricky questions, catches the kinds of misconceptions students try on teachers, and runs without an internet connection. Fewer surprises in the materials you hand out.
A good teaching assistant doesn't just produce content. It pushes back when something is wrong. It notices when a question has been phrased to trip the answerer up. It admits when it isn't sure. This update is about all three.
aime now reasons more carefully before it responds. In practical terms, it catches classic misconceptions — the fraction errors, the place-value slips, the 'subtract the smaller from the larger' habit — and it doesn't fold when a confident student insists 17 × 13 is 222. It works through tricky framing instead of guessing at the answer it thinks you want.
For you, that means two things. The materials you generate are less likely to contain quiet little errors you have to catch in front of the class. And when students use aime themselves, the model is more likely to behave like a patient teacher — guiding, questioning, refusing to short-circuit the thinking — rather than a willing autocomplete that hands over the answer.
It also runs fully offline, so the carefulness travels with you. Your most thoughtful teaching tool doesn't suddenly get sloppier on the day the school internet wobbles.
Trust is built one lesson at a time. This update is one more small deposit.