OpenAI o1 System Card (OpenAI 2024a)

Prior case study of prompting o1 to follow long-term goal and measuring scheming reasoning

Neighborhood — ranked by edge-count

paper

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

OpenAI o1concept0.830
Reasoning model cited as precedent for hidden scratchpad use and scheming capability evaluation
OpenAIinstitute0.741
Lab behind GPT models; shows distinctive post-training signature (denial-before-engagement pattern).
All three OpenAI models show pattern of denying experience first, then describing technical substrate — specific to OpenAI post-trainingfinding0.698
Family voice specific to OpenAI post-training; other RLHF-trained models don't do this
Anthropic Claude 4 system card: two instances in open dialogue develop 'spiritual bliss attractor state' with 'consciousness' emerging in 100% of trialsfinding0.655
Prior empirical observation motivating and converging with the paper's results; self-referential processing between instances producing consciousness claims
Announcing Open-Source SAEs for Llama 3.3 70B and Llama 3.1 8B (Balsam et al., 2025)concept0.651
Goodfire blog post describing SAEs used for Llama models in this study