It’s possible that a French AI start-up quietly started the AI revolution

<p>With much hype and a cleverly produced movie that gave the impression that it was more multimodally adept than it really was, Google unveiled its new Gemini artificial intelligence (AI) model a few days ago.</p>
<p><img decoding=”async” class=”alignnone size-medium wp-image-303857″ src=”https://www.theindiaprint.com/wp-content/uploads/2023/12/theindiaprint.com-its-possible-that-a-french-ai-start-up-quietly-started-the-ai-revolution-mistral-a-750×449.jpg” alt=”theindiaprint.com its possible that a french ai start up quietly started the ai revolution mistral a” width=”750″ height=”449″ title=”It's possible that a French AI start-up quietly started the AI revolution 3″ srcset=”https://www.theindiaprint.com/wp-content/uploads/2023/12/theindiaprint.com-its-possible-that-a-french-ai-start-up-quietly-started-the-ai-revolution-mistral-a-750×449.jpg 750w, https://www.theindiaprint.com/wp-content/uploads/2023/12/theindiaprint.com-its-possible-that-a-french-ai-start-up-quietly-started-the-ai-revolution-mistral-a-1024×613.jpg 1024w, https://www.theindiaprint.com/wp-content/uploads/2023/12/theindiaprint.com-its-possible-that-a-french-ai-start-up-quietly-started-the-ai-revolution-mistral-a-768×459.jpg 768w, https://www.theindiaprint.com/wp-content/uploads/2023/12/theindiaprint.com-its-possible-that-a-french-ai-start-up-quietly-started-the-ai-revolution-mistral-a-150×90.jpg 150w, https://www.theindiaprint.com/wp-content/uploads/2023/12/theindiaprint.com-its-possible-that-a-french-ai-start-up-quietly-started-the-ai-revolution-mistral-a.jpg 1170w” sizes=”(max-width: 750px) 100vw, 750px” /></p>
<p>Mistral, a French AI start-up, has adopted the exact opposite strategy. It first surreptitiously released a download link for its most recent model—which, as it turns out, is really capable—on X, the defunct Twitter platform. This was followed by an official post informing the public of the capabilities of the intriguingly called Mixtral 8x7B. Some of its AI colleagues should be concerned about this model’s open-source nature.</p>
<p>For the Paris-based Mistral, which also saw the start-up earn $415 million, or around €385 million, in Series A investment, it has been an interesting week. Based on approximations, this puts the company’s worth at $2 billion. With the help of the funds, Mistral will be able to expand its commercial line, starting with Mixtral 8x7B.</p>
<p>This model is a follow-up to the September-released Mixtral 7B. And it was just the start. Mistral AI published benchmark comparison results showing that the Mixtral 8x7B outperforms its closest competitors, Meta’s Llama 2 series and OpenAI’s GPT 3.5, by a wide margin. Mixtral 8x7B receives a score of 70.6% on the Massive Multitask Language Understanding benchmark (MMLU), ahead of GPT 3.5 (70%) and Llama 2 (69.9%). Google claimed to be the first AI model to pass this test with a 90% score, at the time of the troika of Gemini model launches.</p>
<p>The other benchmarks follow a similar pattern, with Mixtral 8x7B leading the pack with the ARC Challenge (also known as the Abstract and Reasoning Corpus, a test of common sense reasoning), MBPP (also known as Mostly Basic Pylon Problems, which includes programming problems to be solved), and GSM-8K, which requires the AI models to solve a variety of grade school math word problems. While Mistral’s model did not do well in the multi-turn MT Bench test, GPT-3.5 was victorious in the WinoGrade large-scale dataset issue benchmark, and Llama 2 scored highest in both tests.</p>
<p>Similar to GPT-3.5, Mixtral’s context size is 32,000 tokens per query. The Claude 2.1 large language model (LLM), developed by Google-backed Anthropic, was published last month. It can analyze up to 1,50,000 words in a single prompt, which is equivalent to handling up to 200,000 tokens in a query.</p>
<p>Based on the TruthfulQA/BBQ/BOLD tests for estimating bias and hallucinatory habit, the French start-up claims that the Mixtral 8x7B has more positive sentiment with replies and is more true (73.9% compared to 50.2% of Meta’s Llama 2). Mixtral 8x7B is fluent in English, German, Spanish, Italian, and French, according to Mistral.</p>
<p>The fact that developers may access such a potent and effective AI model in open-source form must alarm companies like Google, Microsoft, OpenAI, and Meta.</p>
<p>Alongside Mixtral 8x7B, we now provide Mixtral 8x7B Instruct. For meticulous instruction following, this model has been optimized using direct preference optimization (DPO) and supervised fine-tuning. With performance similar to GPT3.5, it achieves an MT-Bench score of 8.30, making it the greatest open-source model, according to a statement from Mistral AI.</p>
<p>Mistral AI is unlikely to alter its open-source stance anytime soon. Arthur Mensch, co-founder and CEO of Mistral AI, said in a statement announcing the most recent funding round, “Since the creation of Mistral AI in May, we have been pursuing a clear trajectory: that of creating a European champion with a global vocation in generative artificial intelligence, based on an open, responsible and decentralized approach to technology.”</p>
<p>Beta versions of the Mistral 7B and Mixtral 8x7B models are also being released by Mistral AI. These models will come in three sizes, Mistral-tiny (based on Mistral 7B), Mistral-small (based on the new Mixtral 8x7B), and Mistral-medium, depending on the use case and available processing power.</p>
<p>Our goal has been to construct models that are simple to use and manage by combining the most successful alignment strategies (direct preference optimization, efficient fine-tuning). The business said, “We perform instruction fine-tuning from annotations and pre-train models on data extracted from the open Web.”</p>
<p>Developers may obtain Mistral-small and Mistral-tiny for free. On the other hand, Mistral-medium can only be accessible via a paid application programming interface, or API, that developers and other businesses will plug in.</p>