bf16 breaking VLA fine-tuning sparks debate: fp32 or stochastic rounding already known?
giffmana · x · 2026-11-18
mwulfmeier calls the paper on bf16 breaking VLA fine-tuning one of the month's most interesting; giffmana pushes back that keeping main weights in fp32 or using stochastic rounding has been well established across domains for a while. The exchange highlights low-precision pitfalls when fine-tuning robot policies and known mitigations.
Related event: Paper Claims bf16 Breaks VLA Fine-tuning, Sparking Debate(2 posts)→
More from Embodied
- Concept: a home device to talk to your Grok agents without a phone — nima_owji · 2026-11-18
- GR-BC learns robot skills from failed attempts, hitting ~91% success rate — jiqizhixin · 2026-11-18
- Genesis AI CEO says robots are 6-10 months from the 99.99% success rate needed for mass deployment — davidpattersonx · 2026-11-18
- Lex Sokolin amazed by 'incredible' robot marketing clip — LexSokolin · 2026-11-18
- Neuracore launches Policy Playback: preview robot policy predictions mid-training — stepjamUK · 2026-11-18
- Physical AI panel at MXV Capital AGM features live progress update from Lambda Surgical — ditzikow · 2026-11-18