- Short (3-5s) photorealistic video generation from text prompts. - Stronger priors / Pretraining for Neural Radiance Fields allowing photorealistic one-shot n

- Short (3-5s) photorealistic video generation from text prompts. - Stronger priors / Pretraining for Neural Radiance Fields allowing photorealistic one-shot novel view synthesis.in static scenes. - Scene Flow Fields or equivalent start just-about-working but weaker performance than unstructured video synthesis outside of rigid camera transformations. - Some variant of WebGPT exposed to public, users realize it's already smarter than most people. - Semi-parametric text generation gets terrifyingly good (eg 50x+ improvement for dense parameter equivalent model, not counting lookup table as params)