Vaibhav (VB) Srivastav(@reach_vb):Codex analysed production traffic, improved load balancing, rewrote production GPU kernels and ran hundreds of experiments on its own speculative-decoding model. The kernel improvements reduced end-to-end serving costs by 20%, while speculative decoding improved token-generation efficiency by more than 15%.

Vaibhav (VB) Srivastav

@reach_vb

founder mode @OpenAI | ex @huggingface | F1 fan | Here for @at_sofdog’s wisdom | *opinions my own

加入 June 2017

299 正在关注 55.9K 粉丝

Vaibhav (VB) Srivastav@reach_vb

2026.07.29 21:38

Codex analysed production traffic, improved load balancing, rewrote production GPU kernels and ran hundreds of experiments on its own speculative-decoding model. The kernel improvements reduced end-to-end serving costs by 20%, while speculative decoding improved token-generation efficiency by more than 15%.

显示更多