Research
Findings from our work on data as education — what we measure, what transfers, and why.
The Teaching Signal Beats the Scaling Curve
June 20, 2026 · 1 min read
A study of how curated, instruction-dense corpora move validation loss faster than raw token count — and what that implies for the economics of pre-training.
Measuring What Transfers
June 14, 2026 · 1 min read
Benchmarks reward the skills they name and quietly punish the ones they don’t. We propose a transfer-first evaluation that scores a model on tasks it was never trained to expect.