แบบทดสอบที่เชื่อมโยงกับคำแนะนำ · ยาก ระดับ

เหตุใดวิดีโอ AI จึงได้รับแบบทดสอบผิดทางฟิสิกส์

ทดสอบว่าเหตุใดโมเดลวิดีโอจึงเรียนรู้ลักษณะที่ปรากฏมากกว่าฟิสิกส์ หน่วยความจำที่จำกัดทำให้ความถาวรของวัตถุเสียหายอย่างไร และเหตุใดการคำนวณจึงทำให้คลิปสั้น

คำถาม 1 ของ 8

What does a typical video generator's training objective mainly reward?

ดูตัวอย่างทุกคำถาม8 คำถาม

คำตอบและคำอธิบายจะปรากฏขึ้นเมื่อคุณทำแบบทดสอบ

  1. What does a typical video generator's training objective mainly reward?
    • Obeying conservation of momentum
    • Producing output that resembles real video
    • Matching a built-in physics engine
    • Minimizing the number of frames
  2. Why are rare events like glass shattering especially hard for video models?
    • They are forbidden in training data by law
    • They appear less often in training data and involve complex dynamics
    • They require audio to generate
    • Models cannot render transparent objects at all
  3. The guide describes object permanence failures mainly as what kind of problem?
    • A color-grading problem
    • A memory problem involving limited context
    • A frame-rate problem
    • A problem with the text prompt's language
  4. What is the main reason AI video clips are often only seconds long?
    • Copyright rules limit clip length
    • Attention cost rises steeply as the number of spacetime tokens grows
    • Displays cannot play longer AI clips
    • Longer clips require a different file format
  5. Which statement reflects a misconception about video generators?
    • They may learn some implicit physical regularities
    • They contain a built-in physics engine
    • They can make rare interactions look wrong
    • Their clips are often extended by chaining generations
  6. What did benchmarks such as Physics-IQ suggest?
    • Visual realism reliably indicates physical understanding
    • Visually realistic output does not mean physical understanding
    • All current models pass basic physics tests
    • Physics cannot be tested in video
  7. Why can a small ball's wrong trajectory go uncorrected during training?
    • It takes up few tokens, so the error adds little to the reconstruction loss
    • Balls are removed from training data
    • The model is never shown moving objects
    • Trajectories are handled by a separate audio model
  8. How does Meta's V-JEPA approach differ from pixel-level video generation?
    • It predicts in an abstract representation space instead of pixels
    • It uses only text with no video
    • It relies entirely on a game engine to render output
    • It generates audio instead of video