OpenAI’s CEO, Sam Altman, revealed the company’s newest offering, Sora, on Thursday. With just a text prompt, Sora can produce up to 60 seconds of photorealistic video. Impressive videos that the company claims were made with Sora have been released.
Up until now, ChatGPT and Google’s Gemini—which let users type text into a prompt and receive a text response—have accounted for the majority of user interactions with generative AI. Or maybe you’ve created images using a programme like Midjourney or Dall-E.
For obvious reasons, video is a much more difficult problem to solve. Even so, what OpenAI has so far revealed about Sora is really incredible. The following prompt, “Beautiful, snowy Tokyo city is bustling,” inspired the creation of the video above. The camera pans across the busy city street, following a number of people as they shop at adjacent stalls and take advantage of the lovely snowy weather. Snowflakes and gorgeous sakura petals are swooping in the wind.”
Apart from the videos that the company released, Altman spent the entire afternoon receiving prompts on X and sharing the outcomes. It’s astounding to think that a computer can generate a 60-second photorealistic video just by typing words into a browser window.
However, and this is a crucial point, these are extremely controlled demonstrations. Prompts cannot be typed in by anyone outside of OpenAI to test its capabilities. Because Sora isn’t yet at a point where it consistently yields positive results, journalists who were briefed on the demos were unable to test it out for themselves. Altman even selected examples from user prompts.
Here, there are actually two very distinct things to take into account.
First off, AI is being used in a completely new way. The fact that Sora can achieve these outcomes, even on rare occasions, is impressive and worrisome. According to OpenAI, Sora is being used by safety researchers to establish boundaries and limits. This is likely because the upcoming presidential election season is expected to be the most divisive in history, and Altman does not want to appear before Congress and defend OpenAI’s product’s use in influencing the result.
The other is that I have a lot of questions. How long do these videos take to generate? Hours?
What is Sora being trained with by OpenAI? The project’s co-leader, Bill Peebles, stated in a statement to Wired that “the training data is from content we’ve licenced and also publicly available content.” He did not specify what “publicly available content” entails or if content that is protected by copyright is included.
It is understandable that the majority of people will be preoccupied with concerns about copyright violations and false information, but from the standpoint of OpenAI’s promise, I believe the second question is actually more intriguing. Accurate details are a notorious weakness of generative AI. That is accurate if you type information into ChatGPT or Google’s Gemini, as they will both generate fictitious results with no connection to reality. It applies to both photos and, as of late, videos.
In order to address issues and investigate beneficial uses of this new technology, OpenAI claims to be actively involving legislators, educators, and artists worldwide. “To learn about their concerns and find beneficial applications for this new technology, we will be interacting with policymakers, educators, and artists globally. We are unable to anticipate every positive or negative way that people will use or abuse our technology, even after conducting a great deal of research and testing. Because of this, we think that developing and releasing increasingly safe AI systems over time requires learning from real-world application,” states OpenAI.
