factual
Chronologically trained models as counterfactual factory
Building chronologically trained language models on past research lets you quantify exactly how surprising or improbable a new paper, patent, or product is relative to the model's probability landscape, and—crucially—when an improbable thing turns out true, recompute which other previously improbable ideas have now become probable, enabling funders to take fast 'breaks' toward thousands of newly plausible hypotheses they would not historically pursue.
factualpending
Speaker
James EvansEvidence Quote
“I just got a $20 million grant from the government to basically build what I call chronologically trained language models.”
Source
304 | James Evans on Innovation, Consolidation, and the Science of Science— Sean Carroll's MindscapeCreated: 6/13/2026, 12:24:39 AM
Your notes
Loading notes...