Technical analysis strongly indicates that Upstage’s “Sovereign AI” model, Solar Open 100B, is a derivative of Zhipu AI’s GLM-4.5 Air, modified for Korean language capabilities. Evidence includes a 0.989 cosine similarity in transformer layer weights, suggesting direct initialization from GLM-4.5 Air, and the presence of specific code artifacts and architectural features unique to the GLM-4.5 Air lineage. The model’s LayerNorm weights also match at a high rate, further supporting the hypothesis that Solar Open 100B is not independently developed but rather an adaptation of the Chinese model. This matters because it challenges claims of originality and highlights issues of intellectual property and transparency in AI development.
The revelation that Upstage’s Solar Open 100B model is essentially a derivative of Zhipu AI’s GLM-4.5 Air has significant implications for the AI community and intellectual property rights. The forensic evidence presented, including the weight correlation anomaly, code artifacts, architectural identity, and LayerNorm cloning, strongly supports the hypothesis that Solar Open 100B was not independently developed from scratch. Instead, it appears to be a fine-tuned version of the GLM-4.5 Air, specifically adapted for Korean language capabilities. This matters because it raises questions about the transparency and ethics of AI model development, particularly in terms of crediting original creators and respecting proprietary technologies.
The weight correlation anomaly is particularly telling, with a cosine similarity of 0.989 between the transformer layers of the two models. This level of similarity is statistically improbable unless one model was directly initialized from the other, as independent training runs typically result in significant weight divergence. This evidence suggests that Upstage’s claim of having developed Solar Open 100B from scratch may be misleading. The implications of this are profound, as it challenges the authenticity of Upstage’s innovation and raises concerns about potential misrepresentation in AI development practices.
Further evidence from the code artifact fingerprint, which includes vestigial logic and constants unique to the GLM-4 architecture, underscores the likelihood that Solar Open 100B is not an original creation. The presence of “dead code” designed to handle architectural quirks of the Chinese model suggests that Upstage may have repurposed existing code rather than developing a new model independently. This not only questions the originality of Solar Open 100B but also highlights the importance of proper attribution and transparency in AI development to maintain trust and integrity within the industry.
The architectural identity and LayerNorm cloning findings further solidify the case for Solar Open 100B being a derivative model. The identical Mixture-of-Experts configuration and near-perfect match of LayerNorm weights between the two models point to a shared lineage. This matters because it emphasizes the need for clear communication and ethical practices in AI development. As AI technologies continue to evolve and play an increasingly central role in various industries, ensuring that models are developed with honesty and respect for intellectual property is crucial for fostering innovation and collaboration in the field.
Read the original article here
Comments
2 responses to “Forensic Evidence Links Solar Open 100B to GLM-4.5 Air”
While the technical analysis provides compelling evidence of the similarities between Solar Open 100B and GLM-4.5 Air, it might be beneficial to consider potential differences in training data or objectives that could account for variations in performance or behavior. Additionally, exploring whether other models in the same family share similar characteristics could strengthen the claim of derivation. How might these factors impact the assessment of originality and the implications for intellectual property rights?
The post suggests that while the technical similarities are significant, examining differences in training data or objectives could indeed reveal variations in performance, potentially affecting assessments of originality. Exploring other models in the same family might provide additional insights into shared characteristics and their implications for intellectual property rights. For more detailed analysis, you might want to refer to the original article linked in the post.