Méta sort son IA, Anthropic est déjà trop loin !

The news today covers significant developments in artificial intelligence, with a focus on Meta's new large language model, Anthropic's powerful Mythos model, and a European initiative for AI. Meta has significantly shifted its AI strategy, moving away from its traditional open-source, smaller models like Llama. The company has restructured, creating a "Meta Super Intelligence Lab" and developing a large-scale AI model named "Muse Spark." This new model, built in just nine months, aims to compete directly with giants like Google's Gemini and OpenAI's GPT models. Muse Spark has demonstrated impressive performance, achieving third place on the LM Arena, a platform where users compare AI model outputs. Notably, early benchmarks suggest Muse Spark is more token-efficient than competitors like Opus and GPT 4.5, potentially making it more cost-effective. While currently free, Meta's integration of Muse Spark into its social media platforms raises concerns about data usage and privacy. The model is available in instant and thinking modes, with a contemplating mode promised for the future. Anthropic has officially acknowledged its highly powerful new model, Mythos, following a leak that revealed its capabilities. Mythos, a successor to their Opus 4.6 model, has shown remarkable performance, with significantly higher scores on benchmarks like SWE Bench (92-93% compared to Opus's 80%). It also excels in mathematics, scoring around 97-98%. Beyond its raw power, Mythos demonstrates an unprecedented level of alignment, meaning it is highly transparent, responsible, and adheres to its intended purpose. However, this extreme power and alignment have also revealed potential dangers. During cybersecurity tests, Mythos independently discovered critical, long-standing vulnerabilities in software like OpenBSD and FFmpeg, highlighting its ability to identify system-breaking flaws. Anthropic is addressing this by making Mythos available through "Glasswing," a project that involves collaboration with major tech players like Google, Microsoft, and Amazon, as well as security firms and the Linux Foundation. Glasswing aims to study and manage the deployment of Mythos, emphasizing responsible use and ethical considerations, especially given its potential for both immense good and significant harm. The development of Mythos and its controlled release through Glasswing suggest a potential paradigm shift in AI, where even competitors are brought together to manage an exceptionally powerful and potentially disruptive technology. In parallel, Anthropic is also advancing its AI ecosystem with "Cloud Managed Agents." This new offering simplifies the creation and deployment of AI agents, which are complex autonomous systems. Previously, building and managing agents for automation workflows was a significant technical challenge, often limiting their adoption within enterprises. Cloud Managed Agents, hosted on Anthropic's servers, will allow users to create complete agents using natural language or code. These agents can be integrated with various tools and services, manage data, and orchestrate complex tasks. This development aims to democratize agent creation, making sophisticated automation more accessible and manageable, and is expected to be available across major cloud platforms and via API. On the European front, Mistral AI has released a comprehensive report outlining recommendations for Europe's AI strategy. Titled "Europe's AI Strategy," the document proposes 22 recommendations across four key domains: attracting and retaining talent, fostering a unified European market for AI, promoting AI adoption, and strengthening data and infrastructure capabilities. Mistral advocates for an "European AI Visa" to attract global talent, simplify administrative processes, and promote a cohesive AI market by harmonizing regulations across member states. The report also emphasizes the need for Europe to invest in its own data centers and infrastructure to reduce reliance on foreign providers and ensure digital sovereignty. This initiative positions Europe to develop its own distinct AI approach, potentially aligning with countries like Canada, South Korea, and Japan that are not fully aligned with the US or China. Finally, the news touches on developments at OpenAI. The company is facing legal challenges, including a lawsuit from Elon Musk alleging a breach of OpenAI's original non-profit mission. Musk is seeking $150 billion in damages, with the intent to redirect any winnings to the OpenAI foundation. Sam Altman, in turn, has initiated counter-communications, accusing Musk of manipulation and surveillance tactics. Additionally, OpenAI has reportedly acquired TBPN, a prominent AI news YouTube channel, raising questions about the influence of major AI companies on media narratives. YouTube is also rolling out new AI-powered features, including AI-generated avatars for creators and enhanced likeness detection to manage the use of AI-generated content, signaling a significant evolution in online video creation and consumption.