OpenAI o1 and similar reasoning models (DeepSeek-R1, Claude Extended Thinking) generate a hidden chain-of-thought scratchpad before producing their final answer. The scratchpad contains reasoning steps, intermediate calculations, and self-corrections that are never shown to the user. More compute budget means more thinking steps, which dramatically improves accuracy on hard mathematical and scientific problems - from 18% accuracy for direct answers to 87% for o1 with full compute budget on AIME-level problems.
Animated scratchpad: watch reasoning steps appear one by one in the hidden CoT
Compute budget slider: control how many thinking steps are allowed
Accuracy vs compute scaling curve: see diminishing returns at high budget
Three-way comparison: direct answer vs standard CoT vs o1-style reasoning
Problem difficulty selector: easy arithmetic, medium prime factorization, hard proof
Used in: OpenAI o1/o3, DeepSeek-R1, Anthropic Extended Thinking, QwQ
Part of the EngineersOfAI Interactive 3D - free interactive visualizations covering every major concept in machine learning and AI engineering. Hover any element for a plain-English explanation. No code required.