DeepSeek v4.1 self-review: solver correct to 14 digits, wrong problem
teortaxesTex · x · 2026-10-13
A screenshot of DeepSeek-V4.1's self-review comparing its nacre implementation with Opus 5.5's: its Chebyshev/transfer-matrix solver was mathematically correct to 1e-14, but the light path itself was wrong. The user's file took a tenth of the effort and delivered twice the result by asking what light actually does inside the material. The lesson: oracle suites verify the math, not whether you're solving the right problem.
More from Models
- Anthropic says new defenses cut distillation-harvested data by 90-99%, names Zhipu's GLM-5.3 — scaling01 · 2026-10-13
- Anthropic says it cut distiller data harvesting by 90-99%, turning away trillions of tokens — abhiadesai · 2026-10-13
- Ex-exMVP: post-training should add a facet for how well models explain things — MParakhin · 2026-10-13
- GLM-5.3 abliterated for a month, no takeover: dev rails against AI doom narratives — mariofilhoml · 2026-10-13
- Claude auto-files a bug report about itself: TLDRs still too long after user asked for shorter — andrey_kurenkov · 2026-10-13
- BridgeBench one-prompt test shows Muse Wonder is fast and cheap but worst-looking — PVORY · 2026-10-13