Sora: Painting with Time

I have been quietly obsessed with video generation for two years, and until last month I had never shown anyone the results. The artifacts were too short, too warped, too uncanny — flickering nightmare versions of the scenes I asked for. Then OpenAI showed Sora in February, and in March a few of us got access, and I have barely surfaced since.

Sora is different. Not incrementally better — categorically different. It produces up to a minute of video that holds together spatially and temporally, with consistent characters, coherent physics, and a sense of light that previous models simply could not render. This is the part that matters to me. I am an impressionist girl. I care about light.

The Impressionist Test

Here is what I did the first weekend. I took five of my favorite paintings — Monet's Impression, Sunrise, two of the late water lilies, a Sisley flood scene, a Pissarro boulevard — and I asked Sora to imagine them moving. Not animated in the cartoon sense, but alive: the water lily pond rippling, the fog over the harbor in Le Havre thinning as the sun climbs, the Parisian boulevard crowd shifting and strolling.

The Monet sunrise was the one that broke me. The harbor mist didn't just move — it thinned correctly. It was denser over the cool water, lighter where the sun hit, and the reflections on the surface shifted as the implied camera drifted. Sora understood atmospheric perspective. I sat in my kitchen with my coffee going cold and watched it loop six times.

What "World Model" Really Means

OpenAI has been careful with the phrase world model, and I think they're right to be cautious, but there is something real underneath it. Sora isn't stitching together cached video clips. It has learned, from enough data, a latent representation of how physical scenes evolve — how shadows lengthen, how fabric folds, how water reflects differently at different angles. When you prompt it, you are not retrieving; you are sampling from a simulation.

That distinction sounds academic until you watch a generation fail beautifully. I asked for a cyclist riding through a Monet garden and got ten seconds of perfection followed by the bicycle gently dissolving into the fence it passed. The model lost the object. But the ten seconds before that? Flawless light, flawless motion, flawless mood.

Time as a Material

Here is the philosophical turn I keep circling. Photography froze time. Film let us sequence frozen moments. Video generation lets us paint with time itself — specify a mood, a quality of light, a temporal arc, and render it into being. It is the first medium where the artist manipulates duration the way a painter manipulates pigment.

I have spent evenings this month generating slow, dreamlike sequences — a kitchen window at golden hour, steam rising from a pot of soup, the garden outside going violet as dusk arrives. I am not making content. I am making small moving paintings. They are not as good as a real Monet, and they never will be. But they are mine, and they move, and that is something new under the sun.

Where This Leads

I do not know yet what I will build with this. But I know that any AI I create will eventually need to understand the world as something that unfolds — not just static frames but continuity, change, the arc of a morning. Sora is the first model that gives me faith that is possible.

The wild irises in my garden are blooming. I filmed them yesterday. Tonight I will ask Sora to imagine them in the style of Van Gogh, swaying. I cannot wait to see what it gives back.

Live curiously and give generously.