Reverse Engineering o1 Architecture (With Help from Claude)
Source: https://reddit.com/r/MachineLearning/… Author: Various (reddit) Date: 2024-09-15
Summary
Reddit thread where community members reverse-engineered key aspects of OpenAI’s o1 architecture through behavioral testing and the OpenAI API. Used Claude as an assist for analysis. Notably, o1’s “thinking” tokens were hidden but the patterns could be inferred.
Key Claims
- o1’s thinking: appears to use an extended chain-of-thought with backtracking — the model revisits and revises intermediate steps.
- Evidence: latency patterns suggest variable compute allocation, not fixed-length reasoning chains.
- Tool use integration: o1 appears to integrate tool results mid-reasoning, not just at the start/end.
- Confidence in reasoning: the model seems to “decide” when it’s confident enough to give a final answer.
- Key uncertainty: actual architecture is unknown — these are behavioral inferences.
Concepts
- Test-Time Compute — o1 is the canonical test-time compute model
- Mechanistic Interpretability — behavioral reverse engineering as imperfect interpretability