We collected 10k hours of neuro-language data in our basement (condu.it)

In a groundbreaking effort, researchers have amassed approximately 10,000 hours of neuro-language data from thousands of participants, establishing what is likely the world’s largest dataset of its kind. This initiative aims to train thought-to-text models that decode semantic content from noninvasive neural data. These models demonstrate impressive zero-shot capabilities, predicting semantic phrases from entirely new subjects based solely on neural input, showcasing the potential of decoding brain activity into coherent text. The significance of this dataset lies in its scale and diversity, addressing a prevalent issue in the AI/ML community regarding the lack of large, applicable datasets for training intricate neural decoding models. Participants underwent two-hour sessions of freeform conversations with AI models, facilitated by a custom multimodal headset design that integrates various neural data collection modalities effectively. Innovations included optimizing headset comfort and maximizing data quantity while minimizing noise interference, providing valuable insights for future AI applications. The study's approach to participant engagement and data collection logistics offers critical lessons for conducting research in this rapidly evolving field.