Claude re-checks paper appendix: plain 1-5 ratings (89.6%) match headline SSR figure (90.2%)
scaling01 · x · 2026-11-22
scaling01 quotes Claude's audit of a paper's appendix: plain 1-5 ratings with GPT-4o on text-only concepts score 89.6%, level with the headline SSR figure (90.2%, image stimulus) and above SSR on the same text (82.5%). With Gemini on images, a naive "second LLM assigns the number" baseline edges SSR 92.1% to 90.6%. Conclusion: SSR reliably fixes histogram shape, not which products rank on top.
More from Models
- Inkling 2 Mini tops efficiency charts: 26B looped transformer at $25.80 per correct answer — TheZachMueller · 2026-11-22
- 425MB Qwen3.5-0.8B decision model on CPU: custom 4-bit quant loses just 0.18pp, 2x faster — Danmoreng · 2026-11-22
- Reddit asks: why do OpenAI, Anthropic, and Google models sound almost identical? — retrorays · 2026-11-22
- Scobleizer amplifies claim: Gemini 4 Neon beats Opus 5.5 and costs less — Scobleizer · 2026-11-22
- Julia 1.5 and Julia 2 launch: two models built for decision-making in software — Scobleizer · 2026-11-22
- Inkling Feels Artificially Hyped: First MoE Heavier Than a Same-Size Dense Model — CodeAnguish · 2026-11-22