EX-OPENAI RESEARCHER PREDICTS $100B SPENDING SURGE ON TRAINING DATA

Current large language models are becoming increasingly specialized rather than versatile, excelling at tasks like coding and math while stagnating or regressing in other areas. This trend signals that throwing more compute at existing approaches has diminishing returns. Ho predicts AI labs will need to spend over $100 billion on targeted data collection to address these gaps. The shift reflects a fundamental realization in the industry: quality and specificity of training data may matter more than raw model scale. Ho's departure from OpenAI to focus on specialized training data mirrors broader industry trends. Major AI companies are investing heavily in data curation and synthetic data generation as they hit plateaus with pure scaling strategies. The prediction underscores an emerging bottleneck in AI development—not computing power, but high-quality, domain-specific training data that can push models beyond current limitations.