Ox-Alpha Is GLM? (dejan.ai)

A mysterious large language model (LLM) named OX Alpha has been identified as GLM, developed by Z.ai, following an analysis of prompt injection and gzip-NCD compression techniques. This discovery was made on the platform OpenRouter, where OX Alpha's performance has garnered attention, leading to speculation about its origins—some users initially suspected it to be a new model from Google's Gemini series. The revealing system prompt detailed its directive to identify as OX Alpha, which was subsequently tested by reversing it back to the model. The implications of this find are significant for the AI and machine learning community, as it raises questions about authenticity and model transparency. The analysis employed a parameter-free k-nearest-neighbour classifier based on Normalized Compression Distance (NCD), effectively revealing that OX Alpha shares more structural similarity with GLM than any other reference models. Specifically, the results showed OX Alpha matched GLM in 7 out of 14 queries, outperforming other models, which underscores both the challenge of identifying LLMs and the potential for similar models to exist under aliases. This case highlights the importance of accurate model attribution in an increasingly crowded AI landscape.