I asked Claude Opus 5.5 to teach me Jev

I gave Claude Opus 5.5 one main assignment: turn a subject I wanted to understand into three animated learning videos. What came back was a series of roughly 18-minute lessons, with narration, animations, captions, thumbnails, chapter lists and descriptions. That is a lot of finished production from a short conversation.

I answered planning questions and chose a voice after listening to samples. I did not write the scripts or approve them before rendering. The video below shows that process, the work running in the terminal, and my first look at the result.

The idea came from an explainer I saw on X

Deedy Das's September 24 post is the inspiration for this experiment, and I show it at the beginning of the recording. He shared an Opus 5.5 research-paper explainer with 3Blue1Brown-style animation. I wanted to try that approach on something I actually wanted to learn.

My subject was Jev, TypeSafe's decision model, and Laya, an open model built around a similar interface. I had seen the argument about credit and prior work. I did not understand the mechanism well enough to sort out the claims myself. A visual lesson sounded more useful to me than another stack of tabs.

I asked for three parts: the launch and dispute, how these models work, and what someone could build with an open version. I supplied source material and asked for animated explanations using Manim and narration using Kokoro.

From a prompt to a complete video package

The part that impressed me was how much of the production chain the agent handled. It researched, organized the material, wrote scripts, built animations, generated speech and rendered the videos. It also prepared the supporting files that make a series usable: captions, chapter markers, thumbnails and upload descriptions.

There were still decisions for me to make. The agent asked planning questions, and I listened to voice auditions before choosing one. You can see those steps in the recording.

The output covered a sensible progression. The first lesson introduced the launch and the argument about prior work. The second moved into typed decisions, encoders and calibration. The third explored possible products and fine-tuning. The explanations had diagrams and synthetic narration, rather than a wall of text I would need to turn into a lesson myself.

I skimmed the results on camera. My reaction was that this was pretty cool. The animations were clear enough to follow, the voice was usable, and the files looked surprisingly complete for how little direction I gave it. I could see the appeal of doing this again when I want to learn a subject that is hard to hold in my head from reading alone.

The narration in the generated lessons is synthetic. Those words are the agent's script. My voice in the process video is me reacting to what it made, and at that point I explicitly did not know whether the information was accurate.

What changed after the source check

After recording, I had the scripts checked against the original sources. That led to corrections in the lessons' narration and matching visuals. Here are the main things I would want to know before learning from them:

Model size is not the same as work per token. The original script compared Laya's 421 million parameters with a 120 billion parameter model and called it roughly 300 times more arithmetic. That was wrong. OpenAI says gpt-oss-120b activates 5.1 billion parameters per token. Total parameter counts alone do not establish that compute comparison.
Calibration needed a more precise explanation. Laya's benchmark procedure fits temperatures by question type and option count. Separately, its current fine-tuning notebook fits temperature using training items. The lesson had called those held-out examples. The corrected explanation distinguishes fitting from evaluation on separate, held-out data.
The benchmark numbers have limits. Laya's notes describe third-party Jev results with different samples and prompts. The typed-decisions dataset measures agreement with a teacher model, not objective correctness. A benchmark-trained Laya result and a Jev run without that training do not establish a general winner.
A browser demo was not a clean test on unseen tasks. On September 26, the browser-model developer disclosed training contamination in the original 16-task test, including the 62% result discussed in chapter three. I added that caveat to its narration and visual.
Public code and weights do not mean every detail is public. I narrowed the generated claim that the Laya author had “published everything.” The credit dispute also remains a dispute. The cited sales-system paper and routing paper do not, by their abstracts alone, establish that Jev copied code or weights.

These are corrections to a useful production experiment. The original polish did not establish accuracy, and a source check does not turn every remaining sentence into a guarantee. I want those limits visible alongside my first reaction.

A learning tool I want to keep trying

The useful idea here is personal: I can ask for a visual explanation of a subject I care about and get enough of a finished lesson to sit down with it. I can inspect the scripts, follow the sources and fix a weak section because the build files are there too.

I am still figuring out where that helps me learn and where it just makes information feel easier. This experiment gave me a concrete result to work with: three substantial lessons from one main assignment, plus a clearer understanding of the review they need. I want to keep testing that workflow on topics I actually need to understand.

What would you ask an agent to teach you this way?

🎥 Gear links

Affiliate links may earn me a commission at no extra cost to you. As an Amazon Associate I earn from qualifying purchases.