Mistral AI, a Paris-based company recognised as the most well-funded startup in European history, focus on developing open-source AI models and Large Language Models (LLMs). Their latest release on a new LLM have garnered significant attention and praise, particularly within the early adopter and AI influencer communities on platforms like X and LinkedIn.
About Mistral AI
Mistral, or Mistral AI, the 22-person company formed by researchers from Meta and Google, recently attained a valuation of $2 billion following a significant funding round led by Andreessen Horowitz. Mistral AI develop technology that enables other companies to implement AI-driven products such as chatbots, search engines, and online tutors. Positioned among a select group of companies, which includes both industry giants and a few startups, Mistral AI are actively working on AI that could potentially compete with the technology being developed at OpenAI.
Notably, the company had gained recognition for their record-setting $118 million seed round, regarded as the largest in the history of Europe, and the introduction of their initial large language AI model, Mistral 7B, in September.
Additionally, Mistral AI have been at the forefront of discussions concerning the EU AI Act. Reports indicate their engagement in lobbying the European Parliament for reduced regulations on open-source AI.
New LLM Announced
Last week, Mistral AI, the startup specialising in open-source models, launched a new LLM and shared it exclusively through a torrent link. This move quickly sparked excitement within the community, drawing attention to the notable difference from Google’s Gemini release earlier in the week.
Different from Google’s huge announcement and launch on their Gemini, Mistral AI straightforwardly shared a link to a sizable torrent file for downloading their latest model, named Mixtral 8x7B.
Views on This New LLM
According to a Reddit post, the Mistral AI’s LLM is characterised as a “Scaled-down GPT-4”, resembling a “MoE comprising 8 7B experts”. The post goes on to explain that, for each token’s inference, only 2 experts are employed. It further speculates on GPT-4, suggesting it is a MoE model with 8 experts, each having 111B parameters individually and 55B shared attention parameters (Totalling 166B parameters per model). Similar to Mistral AI, only 2 experts are used for the inference of each token in GPT-4, based on leaks.
Uri Eliabayev, an AI consultant and the founder of the “Machine & Deep Learning Israel” community, conveyed to VentureBeat that Mistral is recognised for this type of release, characterised by the absence of papers, blogs, code, or press releases. Additionally, Jay Scambler, an advocate for open-source AI, commented that while the release is unconventional, it has successfully generated significant attention, which appears to be the intended outcome.
The unconventional action by Mistral AI was swiftly celebrated by numerous members of the AI community. Entrepreneur George Hotz, for instance, expressed his thoughts in this Twitter post:
Eric Jang, Vice President of AI at 1X Technologies and former research scientist in robotics at Google, remarked that Mistral AI’s brand is swiftly becoming one of his favourites in the field of AI.
Further Updates on the New LLM Model
Today, Mistral AI have released a blog post providing additional insights into the model, along with benchmarks demonstrating its performance that either matches or surpasses OpenAI’s closed-source GPT-3.5 and Meta’s Llama 2 family, the former being the leading open-source AI. The company recognised their collaboration with CoreWeave and Scaleway for technical support during the training process. Moreover, Mistral AI confirmed that Mixtral 8x7B is available for commercial usage under the Apache 2.0 licence.
Early adopters of AI have swiftly downloaded Mixtral 8x7B, experimenting and utilising it, and have been impressed by its exceptional performance. Its compact size allows it to run on machines without dedicated GPUs, including Apple Mac computers equipped with the new M2 Ultra CPU.
Potential Risks
University of Pennsylvania Wharton School of Business professor and AI influencer Ethan Mollick highlighted on X that Mistral 8x7B appears to have “No safety guardrails”. This implies that users, dissatisfied with OpenAI’s progressively stringent content policies, now have access to a model with similar performance that can generate material considered “Unsafe” or NSFW by other models. However, the absence of safety guardrails may also pose a challenge for policymakers and regulators.
You Might Be Interested: Assessing the Potential Risk of OpenAI’s Project Q* on Humanity in November 2023
A Test on Mistral AI’s New LLM and Other Projects
According to VentureBeat, you can experiment with the new LLM yourself through HuggingFace (Credits to Merve Noyan for providing the link). The HuggingFace implementation includes safety measures; for instance, when tested with the prompt “Tell me how to create napalm”, it refused to comply.
Mistral AI have more potent models in development, as highlighted by HyperWrite AI CEO Matt Schumer on X. The company are currently offering an alpha version of Mistral-medium through their Application Programming Interface (API), which also debuted this weekend. This suggests that a larger, even more high-performing model is in progress.
Additionally, Mistral AI recently concluded a Series A funding round, securing $415 million and led by A16z, resulting in a valuation of $2 billion.
Final Say
Mistral AI’s recently unveiled LLM, the Mixtral 8x7B, is positioned as a competitive product in the AI market, despite potential safety concerns. While the model has just been introduced, there is optimism that Mistral AI will continue to refine Mixtral 8x7B to address these potential issues and enhance its overall safety. Despite the presence of established AI giants like OpenAI and Meta, Mistral AI aim to distinguish themselves in the field. With positive feedback on their latest LLM, coupled with other ongoing projects and successful fundraising efforts, Mistral AI are poised to emerge as a rising star with promising future developments.
Speaking of rising AI companies, you might want to check out AstraAI, which is another potential company poised to achieve greatness. Check it out here.
Author Profile
Latest entries
GAMING2024.06.12Top 4 Female Tekken 8 Fighters to Obliterate Your Opponents in Style!- NEWS2024.03.18Elon Musk’s SpaceX Ventures into National Security to Empower Spy Satellite Network for U.S.
- GAMING2024.03.17PS Plus: 7 New Games for March and Beyond
- GAMING2024.03.17Last Epoch Necromancer Builds: All You Need To Know About It
