Generate video
Text-to-video is incredible and evolving fast — a quick look, not a deep dive.
The last modality: text-to-video (Sora, Veo, and many others). Karpathy keeps this brief — he doesn't use it in his own work — but flags it as remarkable and moving extremely fast. Worth knowing it exists and roughly where the frontier is.
He shows a comparison tweet: many video models all asked to generate "a tiger in the jungle." All quite good; Veo 2 near state-of-the-art at the time.
Different tools, different styles and quality, all improving fast. Unless you're in a creative field, it's enough to know the capability is here and improving — and to compare a couple when you need it.
The load-bearing points
- Text-to-video exists and is advancing very fast (Sora, Veo, …).
- Quality/style vary by model — compare when you need it.
- Not part of Karpathy's daily workflow; know it's there.
Sample the frontier
If you have access, prompt one text-to-video tool with a simple scene and note the quality and quirks.
Show the point
This modality changes fastest of all — treat anything you see as a snapshot in time.
A light touch — it's impressive and fast-moving, but not something he uses in his own (non-creative) work. Know it exists and where the frontier roughly is.