EpochAI's GPT-6 compute estimate questioned over pretraining-vs-RL split
scaling01 · x · 2026-10-16
scaling01 critiques EpochAI's estimates for GPT-6 compute and parameters: the total training compute seems fine, but the 5.35x ratio of pre-training to RL compute looks off, and the median pretraining duration (90% interval of 42–135 days) is very high. Citing GPT-5.5 (released April 23) as known to be trained on Stargate Abilene, he argues the flaw is that the entire pre + post/RL pipeline started later than EpochAI modeled — not just pre-training.
Related event: EpochAI's GPT-6 Compute Scenario Analysis Sparks Debate(3 posts)→
More from Infra
- Meta picks Apple's MLX as preferred local inference platform; Muse Spark Open runs on two M5 Ultra Mac Studios — awnihannun · 2026-10-16
- Hugging Face scales Indic pretraining with 2.8T tokens and an open data pipeline — Hugging Face Blog · 2026-10-16
- First AA-AgentPerf results: NVIDIA Vera Rubin NVL72 delivers up to 5.7x throughput per MW vs GB300 — ArtificialAnlys · 2026-10-16
- Vera Rubin NVL72 first benchmarks: up to 5.7x throughput per MW vs GB300 — ArtificialAnlys · 2026-10-16
- NVIDIA Vera Rubin NVL72 hits up to 5.7x GB300 throughput per MW on agent benchmark — ArtificialAnlys · 2026-10-16
- 14 open-source LLM observability tools to know in 2026, from Langfuse to Laminar — TheTuringPost · 2026-10-16