Why it matters
A free, public research talk on how to value training data. It gives practitioners a framework for choosing, generating, or transforming data instead of guessing which datasets will help a model generalize.
The tokenmaxxing angle
Indirect: the talk is about training data, not inference bills. But its core idea, spend compute only on data that carries learnable structure, is the same discipline as spending tokens only on context that changes the answer.
From the organizers
The talk starts at 4 PM in the Flatiron Institute's Ingrid Daubechies Auditorium, 162 5th Ave. Registration is required and closes September 26. Wilson is also listed as Research Lead at Perplexity.