Clarifying Opus 5's ARC-AGI Score Details and Compute Scaling Math
imjustnewatai · x · 2026-07-25
The author provides important precision and clarifications regarding the previous analysis of Opus 5's reasoning capabilities:
- ARC-AGI-3 Score Context: Opus 5's 30.16 result on ARC-AGI-3 uses the RHAE evaluation method, meaning it did not literally solve 30% of the game levels.
- Unknowns of J-space: Although Anthropic discovered the internal J-space workspace before RLHF, they explicitly state its relationship to model size remains unknown.
- Compute vs. Parameters: If the 5× annual training compute trend holds for two more years, it implies a 25× increase in raw training FLOPs. However, according to Chinchilla-style scaling, this translates closer to 5× active parameters and 5× training tokens, rather than 25× parameters alone.
Related event: Deep Dive into Opus 5 Hidden Reasoning and ARC-AGI Score(2 posts)→
More from Research
- Oxford launches Philosophy of AI journal, $10,000 prize for best paper — conitzer · 2026-10-27
- Vercel envisions UI future: code, rendering engine and production metrics merge into one loop — shuding · 2026-10-27
- Ex-Vercel engineer Shu Ding: linear approximation gives free layout animations — shuding · 2026-10-27
- Treat designs as points in N-d space: geodesic transitions where no card jumps rows — shuding · 2026-10-27
- Interval arithmetic can prove a UI never fails across infinitely many props — shuding · 2026-10-27
- Vercel Labs compiles React UI into a pure math function, unlocking calculus for interfaces — shuding · 2026-10-27