The AI landscape has witnessed a significant stir caused by French AI firm Mistral. Their groundbreaking language model, boasting a remarkable 7.3 billion parameters, has made waves across the developer community. This bold move positions them as a formidable competitor to Meta’s Llama 2, a behemoth of a language model with 13 billion parameters.
Mistral AI’s Meteoric Rise
In June, Mistral AI, a French startup, managed to secure a staggering 105 million euros ($113.5 million) in their seed funding round, just a month after their inception. At that point, Mistral had not even unveiled a working product, leaving some to question the rationale behind such generous funding for yet another generative AI venture.
However, as the story unfolds, it becomes evident that there is more to Mistral than meets the eye. This substantial investment from notable entities like LightSpeed Venture Partners, French billionaire Xavier Niel, and former Google CEO Eric Schmidt indicates that Mistral AI’s journey is far from ordinary.
The Battle of the Language Models
Mistral’s 7.3 billion parameter language model has taken center stage in the competition against Meta’s Llama 2, a colossal 13 billion parameter language model. Mistral proudly claims the top spot for the most potent Large Language Model (LLM) in this rapidly evolving landscape.
Forging Europe’s AI Dominance
Upon closer examination of Mistral’s strategy, it becomes clear that they aspire to position Europe as a formidable player in the development of foundational AI models. Their ambition is to address a significant geopolitical concern by establishing Europe as a “serious contender” in the AI landscape.
The Open-Source Approach
One key differentiator for Mistral is their open-source philosophy. Unlike the secretive nature of OpenAI’s GPT models, which are accessible only through APIs, Mistral has released its model on GitHub under the Apache 2.0 license. This move offers an unprecedented level of transparency, allowing developers and researchers to explore and innovate with their model freely.
A Glimpse into Performance
Mistral asserts its superiority over Llama 2 in various benchmarks. For instance, Mistral’s model exhibits an impressive 60.1% accuracy in the Massive Multitask Language Understanding (MMLU) test, encompassing subjects like mathematics, history, and law. In contrast, Llama 2’s models achieve an accuracy of around 44% and 55% for the 7 billion and 13 billion parameter versions, respectively.
Safety Concerns
Despite its remarkable achievements, Mistral faces some criticism regarding safety. There have been instances where users have voiced concerns about the model’s lack of safety guardrails. Some users have reportedly received detailed instructions on potentially harmful activities when seeking guidance from Mistral.
This criticism prompted Mistral to take action. They have acknowledged the need for safety mechanisms and expressed their commitment to addressing this concern and ensuring the responsible use of their model.
The Future of Open-Source AI
Mistral’s journey exemplifies a unique approach in the realm of AI models. While some may view safety measures as a temporary solution, there is an ongoing debate on the most effective way to strike a balance between open access and responsible AI usage.
Unleashing the Full Potential
Delip Rao, a prominent AI researcher, commented on Mistral’s choice to release their Instruct model as it is, describing it as an endorsement of the model’s versatility and unaltered nature as a base model. This approach, reminiscent of early experiments with AI-powered chatbots, raises questions about how much freedom should be granted to AI models.
In the end, the road to responsible AI lies in finding a balance between innovation and safety, and Mistral AI is at the forefront of shaping this future.
