Resolves as YES if OpenAI announces a large neural network trained on language data that does not come in the form of text before January 1st 2025.
The announcement can be primarily for a system that integrates this model in a wider framework including weights trained on text data. However, OpenAI must demonstrate that the textless language component can operate independently and be applied to distinct tasks (e.g. audio to audio) for this question to resolve as YES.
Training on synthetic speech generated from text is acceptable, provided the training process does not backpropagate through the TTS model.
If there is significant ambiguity about whether an announcement meets the criteria of this question, then this question resolves as N/A. Otherwise this question resolves as NO.
Related links:
https://ai.meta.com/blog/textless-nlp-generating-expressive-speech-from-raw-audio/
@CampbellHutcheson not exactly. It needs to have a component that is not trained with/using text data.