Skip to content

Pick your style: Denkbaar

Book a conversation

Articles

History

o3 shows the direction of reasoning models

·

OpenAI previewed o3 and invited researchers to take part in safety testing.

A researcher studies patterns on a screen and an assessment sheet.
A test result does not yet prove that a model handles daily work well. Illustration: Denkbaar.

On 20 December 2024, OpenAI gave a first look at o3. The announcement closed its 12 Days of OpenAI series. This was a preview, with early access for safety researchers. At the presentation, OpenAI also showed results on ARC-AGI, a test of difficult reasoning problems. The retrospective lists the o3 preview. A screen shows shape puzzles and several possible steps.

For us, the preview mainly marks the attention given to models that spend more effort on a difficult answer. A test result is different from evidence that a model handles your daily work well. We therefore name no overall winner here. Hands compare two schedules with coloured blocks.

For a small or medium-sized business, the useful question is which task needs extra thought. As a trial, consider comparing two schedules against fixed conditions. First write down what a correct answer looks like and which mistakes matter most.

Then ask a model to perform the same tasks as your existing process. Measure waiting time and review effort as well. Use an announcement as a reason to investigate, without treating a future product as an established part of your business process.

Sources

  1. OpenAI previews o3 (ARC-AGI score)

Topics: model