In the community of Hugging Faceplatform linked tomachine learning e all’Artificial intelligencemysterious files related to a large language model (LLM) of apparently unknown origin.
What revealed the mystery, however, was Arthur MenschCEO of Mistralwho stated that the files in question are the property of his company.
The LLM in question has similar characteristics to Mixtral 8x7b and was described by the company itself as “The highest performing open source LLM available starting January 2024“.
The project was nicknamed Miqu-1-70b, a sort of shorthand for “Mi stral Quantized“. This model is one of the most interesting on the scene as, according to the developers themselves, it could even exceed the potential of GPT-4. How is it possible then that the prototype of this LLM was “stolen” and made public?
Mistral, Mixtral 8x7b and the lanceolate sphid to GPT-4
Mensch himself gave an explanation on the matter, according to which one of the employees (a little too enthusiastic) obtained early access to the model and couldn’t resist publishing an old version of the same on Hugging Face.
At present Mixtral 8x7b represents a sort of evolution of Llama 2 Of Meta, with a pre-training cycle that would have just been completed. Mistral’s CEO himself said he was satisfied with the results achieved, with a possible release (official this time) that could be imminent.
Although in this specific case there were no particular consequences, we must always keep our guard up when it comes to LMM, AI and potential data leaks. Online, in fact, i cybercriminali they are just waiting to get their hands on tools of this kind, to create “personalized” versions that can then be used for illegal purposes, as happened with FraudGPT just a few months ago.