Scaling up foundation models on larger chunks of the internet and expecting emergent capabilities to emerge without domain-specific engineering has run out as a productive research direction because the new capabilities to learn are tacit knowledge, which requires active deployment and feedback loops, not passive training on internet text.