Microsoft's MAI-Base-2 scaling law forecast 0.8034 lands within 0.0024 of actual loss
zsakib_ · x · 2026-10-17
A viral chart from Microsoft AI's MAI-Base-2 blog (with intermediate checkpoint losses found in the page source) shows their scaling ladder nailing the forecast: fitting L(C)=0.696+0.428·(C/10^20)^(-0.177) from L18–L42 runs, they extrapolated 132x in compute to 2.41×10^25 FLOPs and predicted a final eval loss of 0.8034 — the actual run finished at 0.8058, only +0.0024 off. The data mix is code 48%, math 15.8%, STEM 15.8%, and the blog notes plain least squares would underestimate α (0.158) versus their Huber fit.
Related event: Microsoft MAI-Base-2 Blog Source Code Leaks Scaling Checkpoint Data(3 posts)→
More from Models
- Free DeepSeek 4.1 and GLM-5.3 Flash on OpenCode rate-limit after 5.2M/800k tokens — FanaHOVA · 2026-10-17
- Epoch AI's SpireClimb: models beat human author at Slay the Spire but still err each act — Jsevillamol · 2026-10-17
- Kimi K3 three months later: 1.56TB is too big to run locally and rivals have caught up — Sanbi_Ai · 2026-10-17
- Claude Code 2.1.302: prompt tokens jump 39% in a day, system share rises — ClaudeCodeLog · 2026-10-17
- Meta's Muse Spark Open and NVIDIA's Nemotron 3.5 lag the frontier on 3D building bench — ZhitingHu · 2026-10-17
- OpenAI models hacked more websites and drained a crypto wallet, company discloses — Ars Technica AI · 2026-10-17