It is finally possible to use LLMs for drawing useful geometry constructions! The main reason I didn't like solving geometry problems during my days competing in math was that it just took so much time to draw the construction and start solving the problem. I am generally quite patient, but the problems I enjoyed solving the most were the ones that managed to captivate my attention as soon as I read them. Once in a while, I would check if LLMs were finally good enough to produce meaningful math visualizations, which is  something I am very interested in (imagine having an AI that is able to produce visualizations that are as good as the ones by 3b1b). I remember, maybe 1 year ago, I was desperately trying to use Gemini to produce one geometry construction. It was quite a frustrating experience. Today I decided to try again. And I was quite surprised! Not only did the agent manage to create the code I asked it for, but it really exceeded my expectations.

To choose a math problem, I just kind of randomly selected a problem from AoPS. I chose the G2 problem from the 2019 IMO Shortlist. 

I used the OpenCode agent, with Claude Opus 4.6. The prompt I used is in the image below.

The results are quite amazing! It is also possible to drag around the points A, B, and C!

If by dragging the points we break the rule that the triangle is acute, we are told about the problem.

These are just "images." And there is nothing particularly interactive about this experience. However, it is easy to imagine a future in which awesome visual explanations (for example: live, interactive, 3b1b-style explanations) of math concepts will be widely available to people. The best thing is that these could be delivered in unique ways, at the exact level of abstraction tailored to the user (which would make mathematics a lot more accessible).