15 August 2026 · Alex Abramson

Ask Anything: How AI Visual Explanations Work

The pitch is four words: ask anything, see it animate. This is what is happening underneath, using one question as the example.

Say you type: why does a rainbow have a fixed angle?

Step one: decide what to show

People expect this to be the easy step. It is the hard one.

A question like that has a dozen valid explanations at a dozen levels. You could answer with Snell's law and a refractive index table. You could answer with a single ray traced through a droplet. You could answer with the geometry of why the deviation angle has a minimum, which is the actual reason the angle is fixed and not merely the reason there is a rainbow at all.

Choose wrong and you get something technically correct and useless. The model has to work out what the question is really asking, pick a level, then break the explanation into chapters that build. Put everything on screen at once and you have not explained anything; you have produced a diagram somebody needs to be taught to read.

For this question the sequence lands roughly at: one ray entering a droplet, refracting; the same ray reflecting off the back wall and leaving; the colours separating because each wavelength bends by a slightly different amount; and finally many rays at different entry heights, piling up at one angle. The last frame is the answer. The first three are what make it legible.

Step two: write the drawing

This part differs most from what people assume.

Jentoo does not generate a video, and it does not pick from a library of prepared animations. It writes a program. The model outputs real code describing the figure — where the circle is, where the ray enters, what angle it refracts at, where the label goes and what it is attached to — and a sandboxed animation engine executes that code in your browser.

That distinction is the entire reason "ask anything" can be meant literally. A library of clips can only answer questions someone anticipated. A program can be written for a question nobody has ever asked, because the figure is computed rather than retrieved.

It also means the geometry is derived rather than drawn by eye. The refraction angle in that rainbow figure sits where it does because it was calculated, not because it looked about right.

Step three: the narration, and the timing

Every chapter is spoken. That is not a convenience feature; of all the design decisions here it has the most evidence behind it.

Research on multimedia learning has a specific finding here called the modality effect: an explanation delivered as speech alongside a visual generally beats the same explanation delivered as on-screen text alongside the same visual. The reason is bandwidth. Reading text and reading a diagram both compete for your visual attention. Hearing the words leaves your eyes free for the picture.

There is a matching finding on the other side, the redundancy effect, which says narration plus identical on-screen text is worse than narration alone. Two channels carrying the same payload produce interference rather than reinforcement. So the narration is not printed alongside the animation.

The harder requirement is synchronisation. A figure that appears before it is mentioned is a distraction, and one that appears after is a gap. The narration is generated against the finished code, so what is said matches what is on screen at the moment it is said.

Step four: checking it

Generated figures fail in ways generated text does not. Labels land on top of each other. A curve runs off the edge. An arrow points at nothing. None of those are reasoning errors: the explanation can be entirely correct and the picture still unreadable.

So the rendered result gets inspected before it reaches you, and figures that fail get regenerated. It is the least glamorous part of the system and a large fraction of the difference between a demo and something worth using. It is also, honestly, the part still most likely to let something through. We are in beta, and if you get a figure with a label in the wrong place, that is what happened.

What you actually see

About a minute after asking, the first chapter starts drawing — a circle, a ray entering it, the angle marked. A voice explains what refraction is doing. The ray reflects, splits into colours, and the second chapter has begun before you have finished thinking about the first.

By the last chapter there are forty rays entering the droplet at forty different heights, and they visibly bunch at one angle. Nobody has to tell you the conclusion at that point, which is the whole job of a good figure. It makes the answer something you notice rather than something you are told.

The honest limits

Three, stated plainly.

• It can be wrong. The model can misunderstand a question or make a real error, and a confident diagram persuades harder than a confident paragraph. Check anything that matters.
• Watching is not learning. An animation is an excellent way to understand something the first time and a poor substitute for then explaining it back without looking. The techniques that feel like work are the ones that stick.
• Some questions do not want a picture. Asked for a summary of a treaty, Jentoo will draw you something, and text would have served you better.

Within those limits, what it does that nothing else does is straightforward enough: a figure built for your exact question, in the medium the concept actually wants, in about a minute.

Ask it something.

Written by Alex Abramson. Published 15 August 2026.