Mistral CEO Criticizes 'AI Safety Cartel'
Mistral AI's CEO, Arthur Mensch, has challenged what he described as the "AI safety cartel" narrative prevalent in Silicon Valley. His comments suggest a strategic push towards more open innovation and a departure from the risk-management approaches of some competitors. This public stance aligns with the company's recent focus on open-source models and acquisitions.
- Arthur Mensch's "AI safety cartel" critique was made during a keynote at the AI Impact Summit in New Delhi, India, where he argued that a few US companies are using safety rhetoric to create a regulatory moat that stifles competition. He contended that this narrative pressures governments to build policy around a handful of labs, treating others like "unlicensed hobbyists." - Mistral is positioning itself as a vertically integrated, "sovereign" European AI provider to counter US dominance. This strategy is backed by a €1.2 billion investment in a new Swedish data center and the recent acquisition of Koyeb, a serverless cloud platform, to build out its own "Mistral Compute" infrastructure. - The company's open-source focus is a key part of its strategy, with models like Mixtral 8x7B released under the permissive Apache 2.0 license. This model utilizes a Sparse Mixture-of-Experts (SMoE) architecture, which allows it to outperform larger models like Llama 2 70B with 6x faster inference speeds. - Financially, Mistral has raised over $3 billion to date, including a recent €1.7 billion Series C round led by semiconductor equipment manufacturer ASML, valuing the company at €11.7 billion. This partnership with ASML includes an 11% stake and aims to integrate AI into semiconductor manufacturing. - The term "AI safety cartel" points to differing philosophies on managing AI risk. While OpenAI primarily uses Reinforcement Learning from Human Feedback (RLHF), Anthropic employs "Constitutional AI," training models with a set of explicit principles. Mensch argues that true safety comes from transparency and the global community red-teaming open models. - Mensch's comments were delivered to an audience of global policymakers and regulators, framing the debate as one of economic sovereignty versus dependence on foreign technology. He urged nations to invest in their own AI infrastructure to avoid a world where "three or four enormous companies actually own the deployment and the making of AI."
