SORA is finally here

Day 346 / 366

Back in February, OpenAI announced their text-to-video model SORA, with amazingly real-looking samples. The demo was great, and I couldn't wait to try my hands on it. But like everything else AI, this one also never actually came out for the general public. This changed a few days ago when they finally released it. And the results have been underwhelming.

For starters, even though I had a Plus subscription, I was not able to access SORA and try it out myself. Seems like they disabled new account creation.

But I did find some examples of people using it on Twitter. And the results were hilarious.

Here’s a video SORA generated of people playing cricket -

Neither is this real-looking nor does it get any of the physics right. In the case of LLM’s, we have this thing called emergent behaviours. These are things that the model learns on its own, even though it is not specifically trained for it.

For instance, most of the GPT models know how to write code. They were just trained on text from the internet, and they just happened to somehow learn how to write code from that.

The claims that many were making about SORA were similar, that just by getting trained on videos from the internet, the model will learn basic physics. But that is clearly not the case. Here is another video of a gymnast that is AI-generated —

SORA has been a disappointment, let’s hope that this is just the beginning and that we will get better models in the future.