Ex-OpenAI Researcher Bets $100B on Training Data

Andrew Ho spent eight months at OpenAI before deciding the company's core bet, that scaling large language models will eventually produce systems that generalize across tasks, doesn't hold up. He's now left to start a company built on a different premise: the bottleneck isn't compute or model size, it's data. Specifically, the kind of data that captures how humans actually do skilled work, from bioinformatics to routine lab procedures, none of which shows up in the text scraped from the internet that trains today's models.

Ho's new venture is building specialized datasets meant to fill that gap, and he's not shy about the price tag. He expects AI labs will collectively need to spend upward of $100 billion on targeted data collection to get anywhere close to the reliability they're promising customers. That's a striking claim from someone who spent less than a year inside one of the labs setting the pace for the industry.

His skepticism isn't isolated. Researchers at the University of Cambridge and Google DeepMind have pointed to similar trends, noting that current systems seem to be drifting toward narrow specialization rather than the broad, flexible intelligence labs keep promising.