Mistral, an artificial intelligence (AI) startup, has launched Mistral 7B, its open-source language model. This model can be used by anyone free of charge and in a variety of ways. Worth 240 million euros, Mistral made available in June a preliminary version of its free and generative language model, which does not seem to have any special moderation mechanisms. The company emphasizes the superiority of its product over competing solutions.
Mistral 7B has the capacity to process and generate text faster than larger proprietary models while costing a small fraction of those models. Currently, the language model can already be used for a variety of tasks such as summarizing, structuring and answering questions.
Mistral 7B has been released under the Apache 2.0 license, which essentially means that it can be used in any context without any restrictions. Likewise, there are no limits on the content that can be generated by the AI model. Security expert Paul Röttger has noted that the model is capable of executing instructions for various criminal activities.
Another point of criticism from Röttger relates to the publication of the model via a magnet link using torrent by Mistral. This has the consequence that, already after its distribution on many systems, the model cannot be withdrawn and supplemented with moderation elements afterwards.
Röttger explains, “My biggest problem with Mistral’s release is that security was not analyzed or mentioned in their public messages.”
Mistral recently confirmed that Mistral 7B does not have a moderation mechanism. It is worth mentioning that this information was originally missing from the initial release and was added after the fact.
There is heated debate as to whether it makes sense to release an open language model without moderation. Some users call for limits on the use of the model due to security considerations to prevent criminal activity. Others argue for unrestricted availability and accessibility of the language models, without any censorship or control.
