Sora2 is way too advanced.
It can predict ChatGPT’s output and render HTML?!
When asked to simulate “sending a message to ChatGPT,” it not only generated visuals but also performed a Q&A “interaction.”
First, it crafted a prompt: Write a playful haiku about a cat staring out the window.
Then, mimicking ChatGPT’s style, it delivered an audio response: Whiskers pressed to glass. Birds gossip beyond the pain. Tail flicks. Daydreams fly. Tail flicks. Daydreams fly.“
The entire response was delivered in ChatGPT’s mechanical female voice, with the haiku syllables perfectly aligned.
This video scene + LLM reasoning demo amazed netizens, with some even stating, ”Sora2 blurs the line between video generation and interactive AI.”
In fact, Sora2 isn’t just capable of predicting ChatGPT-style reasoning responses—it can also render HTML.
Passed the glass refraction test
Someone else had Sora2 render pasted HTML code, yielding the following result:
And here’s how that same code renders in a real browser:
It’s not just similar—it’s identical.
Seems Sora2 isn’t just a video generator—it can double as a browser.
Beyond simulating interactions and understanding code, Sora2 even demonstrates a grasp of physics.
When prompted:
Place a vertical piece of paper on a table with two arrows drawn on it—one below the other, both pointing in the same direction. Position a large glass cup in front of the arrows so they’re visible through the cup. Then fill the cup with liquid.
The result was:
What?? It proactively demonstrated glass refraction without explicit instructions to flip the camera angle.
Netizens reacted with: insane.
Perhaps Sora2 knows it’s being “tested.” (doge)
Additionally, users discovered that without detailed prompts, Sora2 accurately recreated an element from a side quest in the game Cyberpunk.
Prompt:
Generate gameplay footage of using the Dragon Tank and Panum in Cyberpunk 2077.
Key elements like map locations, biomes, terrain, vehicle designs, and gang names were all remembered. Only two minor details were incorrect:
One was depicting the Dragon Tank’s movement as wheeled (it should actually hover).
The other was an incorrect placement of the Panum (it should be inside the tank, not on the turret).
The author believes that while video tutorials for this side quest exist online, it remains a side quest. Sora2’s ability to accurately extract and integrate key details from vast information reflects a certain level of intelligence and information processing capability.
Regarding these impressive performances, many speculate that Sora2 might be trained on large language models (LLMs).
Perhaps we haven’t yet uncovered Sora2’s other “special abilities.”
What are your thoughts? What fun things have you generated with Sora2?