Preliminary results suggest V-JEPA representations let the system tell whether a video is physically possible or impossible.

The claim is hedged as preliminary. True if, by 2024-03-08, Meta/FAIR had preliminary results (internal or published, such as the V-JEPA paper or blog post, or follow-up work) showing that V-JEPA representations can be used to tell physically possible videos from physically impossible ones (for example, a surprise signal on violation-of-expectation tests). Mostly true if the results showed this only partly, for some physical properties only, while still supporting the preliminary claim. Mixed if the results were ambiguous. Mostly false if the evidence came only later and no results existed at the time, or if they showed only general video understanding rather than physical possibility. False if the evidence shows V-JEPA representations failed to tell possible from impossible videos.

Set before the research began.

1. E1 How One AI Model Creates a Physical Intuition of Its Environment | Quanta Magazine Quanta Magazine · 3 October 2025

   “On a test called IntPhys, which requires AI models to identify if the actions happening in a video are physically plausible or implausible, V-JEPA was nearly 98% accurate.”

2. E2 How One AI Model Creates a Physical Intuition of Its Environment | Quanta Magazine Quanta Magazine · 3 October 2025

   “The team found that the prediction error shot up when the future frames contained physically impossible events.”

3. E3 Intuitive physics understanding emerges from self-supervised pretraining on natural videos arXiv (rendered via ar5iv)

   “we find that video prediction models trained to predict outcomes in a learned representation space demonstrate an understanding of various intuitive physics properties, such as object permanence and shape consistency.”

4. E4 AI at Meta Blog Meta AI · 15 February 2024 Own record

   “We’re releasing the Video Joint Embedding Predictive Architecture (V-JEPA) model, a crucial step in advancing machine intelligence with a more grounded understanding of the world.”

The evidence shows that V-JEPA-type models can tell physically possible videos from impossible ones [E1, E2, E3], but these items are from 2025 or are undated. The only item dated before 2024-03-08 [E4] does not mention any physics results, so the evidence does not show whether such results existed when the claim was made.