← Back to People
Authored Library of Layers transformer architecture allowing dynamic reordering of transformer blocks and discovered quantization time-bomb loss spikes in BF16 training.
2
Mentions
2
Articles
1
Outlets
Articles
Most recent firstMentioned In
Referenced in coverage
Delayed Loss Spikes in Quantization
XneutralNov 4, 2026
“Discovered delayed loss spikes driven by quantization error in the LM head during BF16 training.”
Library of Layers Transformer Architecture
XpositiveNov 4, 2026
“Authored Library of Layers transformer architecture allowing dynamic reordering of transformer blocks and discovered quantization time-bomb loss spikes in BF16 training.”