Resolves as YES if there is strong evidence that OpenAI has added and publicly released (true multimodal) image generation capabilities to its o1/o3/... series models before January 1st 2026.
Note that if OpenAI rolls its o* and GPT-* models into a common neural network backbone, then this question resolves as YES only if the image tokens are produced by this model. Using an intermediate text representation and dispatching this to a surrogate model does not count.
Access to these capabilities must be available in at least 4 of the following countries: USA, UK, France, Japan, India, Canada, Australia, Germany, Italy before the specified date.
1,000
@Ernie yeah, I got the impression with their AMA that they were saying they had some good proposals internally of what to scale up, but for compute they had multiple higher priority items. That being said, I think they really do care about image generation as a problem, and that it's highly relevant to demonstrating that their RL approach generalises to non-text tokens.