TL;DR
Get the latest gadgets delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Rio de Janeiro’s claimed original large language model appears to be a merge of existing models Nex and Qwen. Evidence shows it is not independently trained. This raises concerns about transparency and local AI development claims.
Recent technical analysis indicates that Rio de Janeiro’s locally developed large language model (LLM), announced as an original creation by IplanRIO, is in fact a weighted merge of two existing models, Nex and Qwen, rather than a fully independent training effort.
Hacker News researchers examined Rio’s ‘Rio 3.5’ model, presented as a 397-billion-parameter model trained by IplanRIO. Their analysis found that its weights are a direct element-wise blend of Nex and Qwen models, with approximately 60% Nex and 40% Qwen, across all network layers. They observed that when the model’s system prompt is removed, it identifies itself as ‘Nex, from Nex-AGI’ 79% of the time, and as ‘Rio’ only 0%.
Further, every tensor in the model matches the weighted combination of the two source models to thousands of standard deviations, with no evidence of independent training or fine-tuning. The researchers conclude that Rio’s model is essentially a composite of pre-existing models, not a new, trained-from-scratch model as claimed.
Implications for Local AI Development Transparency
This development raises questions about the transparency and authenticity of local AI initiatives, especially when claims of original training are contradicted by technical evidence. It highlights the potential for misrepresentation in AI model development efforts and underscores the importance of open verification.
For stakeholders in AI and government, this could impact trust in local AI projects and influence future funding and policy decisions. It also emphasizes the need for independent audits and disclosures in AI development claims.
large language model training kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Rio’s AI Claims and Model Analysis
Rio de Janeiro announced the development of an ‘original’ large language model, called Rio 3.5, purportedly trained by IplanRIO. The project was positioned as a homegrown effort to advance local AI capabilities. However, recent analysis by independent researchers on Hacker News examined the model’s internal weights and behavior, revealing it to be a blend of two existing models, Nex and Qwen.
This finding suggests that the model was not independently trained from scratch, as initially claimed, but rather assembled by merging pre-trained weights. The analysis provides a timeline of the development claims versus the technical evidence, raising questions about the accuracy of the original announcement.
“Every weight tensor in Rio is, to thousands of standard deviations, the same 0.6/0.4 blend of Nex and Qwen across all layers.”
— an anonymous researcher
As an affiliate, we earn on qualifying purchases.
Unconfirmed Aspects of Rio’s Model Development
It is not yet clear whether any additional fine-tuning or modifications were performed after the merge, or if the model was presented as fully original to obscure its true nature. The developers have not publicly addressed these findings, and official statements remain absent.
neural network analysis software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Verification and Transparency
Further independent analysis and potential disclosures from IplanRIO or authorities may clarify whether additional training or fine-tuning occurred. Future audits could determine if the model’s claimed originality is accurate or if similar cases of model merging are widespread in local AI projects.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is Rio’s ‘homegrown’ model truly original?
Current evidence indicates it is a weighted merge of Nex and Qwen, not an independently trained model, contradicting initial claims.
Why does this matter for AI development in Rio?
It raises concerns about transparency, authenticity of local AI claims, and trust in publicly announced projects.
Could the model have been fine-tuned after merging?
This remains unconfirmed; the analysis focused on weight similarities, and no evidence of additional training has been publicly disclosed.
Will this affect future AI projects in Rio?
Potentially, as it may lead to increased scrutiny and calls for transparent development practices in local AI initiatives.
What is the significance of this finding for AI research?
It highlights the importance of verifying claims through technical analysis and the risks of misrepresenting model origins.
Source: Hacker News
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
