rgreenblatt/which_claude_is_k3

This repository investigates which large language model Moonshot's Kimi K3 was likely trained on or influenced by. The author says the project uses identity probes, tests that compare model responses, applied to Kimi K3, related models in its lineage, some open source models used as a baseline, and reference calls to Claude and GPT models. The most striking detail in the README is a direct admission: every part of this project, including the experiment design, the code, the data collection, the analysis, and the final write up, was produced by an AI agent, described as Opus 4.8, working without human review of the results or claims. The author explicitly warns that the findings may contain errors and should not be trusted as reliable without someone independently checking the work. The repository points readers to a full report file for the detailed findings, and to a separate cleaned up folder that contains a curated, reproducible version of the code and results, including figures that can be regenerated offline from the results that are already included. The root of the repository is described as the original working tree used during development, with a progress log file that records the full history of how the project evolved. The curated folder is presented as the polished deliverable meant for people who just want to review or rerun the analysis without wading through the development history. Given the small size of this project, it does not describe installation steps, dependencies, or how to run the code, so those details are not available. The repository is written in Python and has not accumulated much attention yet, with only a couple of stars.