A 2019 paper coauthored by Harvard's Boaz Barak showed that double descent can occur not only as models gain parameters, but also when they receive more training data or are trained for longer.
Notes on verification
Directly confirmed by the paper's abstract (arXiv:1912.02292), corroborating OpenAI blog post, and independent confirmation of Barak's Harvard affiliation. [tier=gold indep_score=0.9 clusters=4 claim_tier=notable]
Sources
- Large language models can do jaw-dropping things. But nobody ... (seed:technology_and_ai)
- https://arxiv.org/abs/1912.02292 (corroboration)
- https://openai.com/index/deep-double-descent/ (corroboration)
- https://windowsontheory.org/2019/12/05/deep-double-descent/ (corroboration)
- https://quantum.harvard.edu/boaz-barak (corroboration)