Welcome back to the AI For Humans Newsletter!

Each week on AI For Humans, we spend at least a little time diving into new AI video tools and trying to show people how easy it’s getting to create at least passable shots for all sorts of productions.

But yesterday, a new technical paper dropped that might have just yanked our “prompt-to-Hollywood” timeline a lot closer.

This new paper shows how a team of researchers managed to get an AI model to generate entire minute-long videos (with multiple shots and scenes!) just from a single text prompt. Normally, AI video models start strong and fall apart fast, weird warping, broken continuity, characters morphing into new people mid-shot. But this one stays coherent across time, which is a huge deal. It even learned a kind of “memory,” so characters and visual styles stick around from scene to scene.

The secret sauce here is something called “Test-Time Training”. Basically, the model keeps learning even while it’s generating. It’s a bit like a director making edits on the fly, adjusting the model’s internal state to keep things consistent and on-theme. They tested it by training on clips of Tom & Jerry (yes, seriously), and the results are surprisingly cinematic. It’s not perfect yet but it’s way closer than most other models we’ve seen.

And the prompts? Long. Way longer than you'd expect. But the model follows them, shot by shot, across a full minute of coherent video.

What makes this cool isn’t just the better output, it’s what it means for future tools & use cases in both Hollywood & independent productions.

Imagine Pixar-level storytelling but without the years-long production grind. That’s the kind of leap this tech hints at. Studio-quality vibes, solo creator speed.

More on Thursday!

-Gavin (and Kevin)

We 💛 This - New 4o Image Generation Formats

We posted a tweet that has gone viral around a fun new prompt that we found on the Sora homepage. It’s spurred on a bunch of people to try it and it’s a good reminder that while OpenAI’s 4o Image Generation isn’t as powerful as it was on launch, there’s still some fun stuff you can try.

We found this originally browsing the Sora homepage (and created the Tony Hawk version you see in the tweet above). Browsing others’ creations is an amazing way to see what the new model is capable of & what OpenAI is still allowing it to do. Like create insanely yoked versions of the Superbad cast. For that, we’re very sorry.