Club Benefits
You are now subscribed
Your newsletter sign-up was successful
Want to add more newsletters?
Daily (Mon-Sun)
Tom's Guide Daily
Sign up to get the latest updates on all of your favorite content! From cutting-edge tech news and the hottest streaming buzz to unbeatable deals on the best products and in-depth reviews, we’ve got you covered.
Weekly on Thursday
Tom's AI Guide
Be AI savvy with your weekly newsletter summing up all the biggest AI news you need to know. Plus, analysis from our AI editor and tips on how to use the latest AI tools!
Weekly on Friday
Tom's iGuide
Unlock the vast world of Apple news straight to your inbox. With coverage on everything from exciting product launches to essential software updates, this is your go-to source for the latest updates on all the best Apple content.
Weekly on Monday
Tom's Streaming Guide
Our weekly newsletter is expertly crafted to immerse you in the world of streaming. Stay updated on the latest releases and our top recommendations across your favorite streaming platforms.
Join the club
Get full access to premium articles, exclusive features and a growing list of member rewards.
AI video generators can’t understand the laws of physics solely by watching videos, scientists have found.
Coming hot on the heels of chatbots and image generators, AI video generators like Sora and Runway have already been delivering impressive results. But a team of scientists from Bytedance Research, Tsinghua University, and Technion were curious to learn if such models could discover physical laws from visual data without any additional human input.
The three fundamental physical laws for simulation they chose to study were the uniform linear motion of a ball, the perfectly elastic collision between two balls, and the parabolic motion of a ball.
Based on the team's pre-print paper, it turned out that while the shapes acted as they should for simulations based on the data they were trained on, they failed to act properly in new, unforeseen scenarios. At best, the models tried to mimic the closest training example they could find.
During the course of their experiments, the scientists also observed that the video generator often changed one shape into another (e.g. a square randomly turns into a ball) or made other nonsensical adjustments. The model's priorities appeared to follow a clear hierarchy, with color holding the highest importance, followed by size, and then velocity. Shape received the least emphasis.
Have they found a solution?
“It is challenging to determine whether a video model has learned a law instead of merely memorizing the data,” the researchers said. They explained that since the model’s internal knowledge is inaccessible, they could only infer the model’s understanding by examining its predictions on unseen scenarios.
“Our in-depth analysis suggests that video model generalization relies more on referencing similar training examples rather than learning universal rules,” they said, highlighting this happens regardless of the amount of data a model trains on.
Have they found a solution? Not yet, lead author Bingyi Kang wrote on X. “Actually, this is probably the mission of the whole AI community,” he added.
More from Tom's Guide
- 5 Best AI video generators — tested and compared
- AI glossary: all the key terms explained including LLM, models, tokens and chatbots
- Meet Mochi-1 — the latest free and open-source AI video model
