Two Prompts: Claude Opus 5.5 Turns a Neglected Meta Paper into an Explainer Video

Anthropic's Claude Opus 5.5 used two prompts to turn a Meta FAIR paper into an explainer video. The poster, Tanishq Mathew Abraham, said that after watching the video, he basically understood how the paper's method JEM works.

The generated result is as follows.

The video is about 'Scalable Patch-Level Self-Supervised Learning', submitted to arXiv on October 7, 2026, with authors including Maximilian Seitzer, Maxime Oquab, Oriane Siméoni, and others, all from Meta FAIR. X user @iScienceLuvr said in the tweet that this paper 'flew completely under the radar' and almost no one noticed it.

The paper proposes JEM, a student-teacher self-supervised learning framework. Without human annotations, the model learns representations directly from images. JEM works by splitting different views of the same image into patches, encoding them separately with a teacher network and a student network, and then aligning patch representations across views. One type preserves information, the other preserves structure, preventing representation collapse.

The paper has three hard metrics:

• Training is stable from 300M to 7B parameters; it is the first method to align latent-space patch representations at 7B scale.
• Strong performance on global and dense tasks; consistently outperforms DINOv2 on segmentation benchmarks.
• At 7B parameters, it surpasses DINOv3 on panoptic segmentation, with about 1/12 of DINOv3's training data and no refinement stage.

DINOv2 and DINOv3 are both Meta's own visual self-supervised models and are de facto baselines in this field. JEM's results would normally be enough to trigger a round of reposts, but it had been published for two weeks with almost no discussion. That was until Tanishq retweeted this post with a video attached.

Tanishq posted the prompts he used. First prompt:

Make a video explaining how JEM works (https://arxiv.org/abs/2610.10013)

Make sure it's in depth but also explains it from scratch and first principles without it being too complicated (I don't understand any of the information theoretic math for example, it confuses me)

The gist is: make a video explaining how JEM works, go deep, start from first principles, and don't make it too complicated. He specifically added: 'I don't understand the information-theoretic math; that stuff confuses me.'

The result was a 16-minute video that did not involve formulas. Second prompt:

This is incredible! The only thing missing is it doesn't walk through the actual math/equations for the losses. You just need to explain how your intuitive description of the losses connect with the math. Please let me know what parts of the video you change to make this clear.

The gist is: the video is great, but it is missing one piece—it needs to connect the intuitive explanation with the math of the loss functions.

The result was a 25-minute version. Tanishq said it was just these two prompts, without using any third-party tools.

The comment section started with technical Q&A.

One user asked: The DINO series uses a bunch of operations like centering to prevent collapse; what does JEM rely on? Tanishq replied: an entropy-based regularization term.

The phrase '12x less data' in the abstract is also ambiguous. Some felt the subject could be read as either JEM or DINOv3. After checking, a user clarified: the subject is JEM. At 7B parameters for panoptic segmentation, the data is about 1/12 of DINOv3's.

Another comment discussed engineering cost, saying DINO's recipe contains many stabilization tricks, and most teams only download official weights and never retrain on their own data. If JEM is truly simple enough, the training cost for domain-specific backbones would drop significantly.

Others were not convinced. ews said that AI-generated videos may look completely self-consistent, but no one can guarantee that they correspond one-to-one with the paper. It is like giving an LLM an English mathematical proof and having it generate a corresponding Lean formal proof, but the two are not necessarily equivalent. Sparsh Jain also asked: Has anyone compared the video with the paper item by item? Did the model get anything wrong?

But at least the barrier to reading papers is being flattened by tools. People who don't understand information-theoretic math can also use two prompts to generate a video for themselves that explains a paper clearly. As for whether the formula details are reliable, the paper is below and can be checked.

arXiv:2610.10013