Sam Altman, in a discussion with Nicholas Thompson of The Atlantic, hinted at the possibility of training AI models entirely on synthetic data. During the podcast, Thompson highlighted the growing prevalence of AI-generated content and suggested that future AI models might be trained using such data. Altman, when questioned about training models solely on synthetic data, hesitated before implying a confirmation, stating that reasoning skills could be developed with synthetic data alone. Altman elaborated that while AI could potentially excel in tasks like mathematics without human data, understanding human values would likely require exposure to human culture. He addressed concerns about the degradation of information through repeated AI output consumption, likening it to the "mad cow disease" analogy, but maintained that teaching AI to perform calculations doesn't necessitate human data, whereas understanding human nuances does.