Chart claims 12.3x training FLOPs efficiency vs DeepSeek-V4-Pro-Base, sparking scaling debate
scaling01 · x · 2026-10-16
scaling01 shares a chart showing an MAI base model achieving better bits-per-byte on held-out code with 12.3x better training FLOPs efficiency than DeepSeek-V4-Pro-Base, plotted along a scaling ladder. He jokes about "Jevons paradox in Chinese" and asks whether Chinese neolabs are already crashing — a snarky take on what frontier training-efficiency gains mean for Chinese labs chasing the curve.
More from Models
- Gemini 4 Argon rolls out globally in Google Search's AI Mode for Pro/Ultra subscribers — OnlineInference · 2026-10-16
- Dev Builds Haunted Escape Room to Test Reflection AI's Beam, Solves It in 30 Seconds — wandb · 2026-10-16
- Claude Opus 5.5 fast mode exits preview: 2.5x faster at $8/$40 per million tokens — ClaudeDevs · 2026-10-16
- NVIDIA ships Nemotron 3.5 Bolt: 74B-A9B model in 24GB of memory — NVIDIAAI · 2026-10-16
- User's Opus 5.5 High Blender 3D experiment crashes their Mac with hilarious results — imjustnewatai · 2026-10-16
- Apollo Research: GLM-5.3 Spontaneously Reasons About How Its Actions Shape Future Training — jd_pressman · 2026-10-16