Leaked Google 'Mathematica' Model's Raw Chain-of-Thought Screams 'OH MY GOD' at Elegant Proof
新智元 · wechat · 2026-09-20
Leaked internal Google evaluation screenshots reveal 'Mathematica,' a math-specialized variant of the unreleased DeepThinkV3. While solving a hard Diophantine equation, its unfiltered raw chain-of-thought erupts in all-caps exclamations and muses on 'the absolute beauty of this equation.'
Key configuration details:
- 65,536-token output cap (vs. typical 4K-8K), enabling massive multi-step derivations in one response
- 1.04M-token input context, inherited from Gemini's long-context architecture
- Default Temperature=1 — unusually high for math, apparently to encourage intuitive leaps
- Endpoint tagged deepthink-mathematica-tf-raw-thoughts ('tf' = Teamfood internal testing, status UNSTABLEEXPERIMENTAL)
The piece explains the 'emotional' CoT as a mechanism artifact: raw thoughts skip RLHF polishing, human math-forum exuberance in training data binds exclamations to breakthrough moments, and Process Reward Model spikes at elegant simplifications project as excited punctuation — an algorithmic 'eureka.' It closes with reflections on mathematical beauty and creativity as signals on the road to AGI.
More from Models
- Dev compares coding models building a coop game: V4.1 outshines Astra's 'pathetic' default taste — teortaxesTex · 2026-09-21
- TypeSafe's Jev returns typed decisions with probabilities, not text — here's where it fits in agent loops — prakersh · 2026-09-21
- DeepSeek-V4.1-Flash redesigns the Transformer for agents, cutting KV cache to 890 bytes/token — AndLukyane · 2026-09-21
- Grok 4.7 briefly spotted on opencode zen, hinting at imminent xAI release — socialwithaayan · 2026-09-21
- NIST's CAISI publishes assessment of Z.ai's GLM-5.3 cyber capabilities — 233C · 2026-09-21
- Grok 4.7 briefly appears on OpenCode Zen before deletion, hinting at imminent release — kimmonismus · 2026-09-21