Introduction
February 2025
Major events
February was the month of the Paris AI Action Summit, and of the competing ideas of AI it put on one stage. France pitched its nuclear-powered electricity and announced €109 billion in prospective private investment; the European Commission launched InvestAI to mobilize €200 billion, including funds for AI “gigafactories”; European companies formed an AI Champions initiative. US Vice President JD Vance used the same summit to argue against heavy regulation. Days earlier, Google had dropped its pledge not to apply AI to weapons or surveillance.
The companies kept shipping. Google made Gemini 2.0 Flash generally available, Mistral took Le Chat to phones, Adobe and YouTube put generated video into creative tools, and xAI unveiled Grok 3 with reasoning models. Control was contested too: OpenAI’s board rejected a $97.4 billion offer from an Elon Musk-led consortium, publishers sued Cohere over their content, and South Korea’s privacy regulator saw DeepSeek’s app pulled from local stores while it addressed the country’s data law.
Patterns and trends
Two approaches to governing AI stood side by side: Europe paired large investment programs with the AI Act, whose first obligations had applied on February 2, while Washington argued that regulation would hold the industry back. Companies, meanwhile, moved to own more of what they depended on. OpenAI was reported to be finishing its first chip design, and Figure AI left its collaboration with OpenAI to build its own robot models. Research looked for cheaper routes to better models: Anthropic’s Constitutional Classifiers resisted jailbreaks in a public challenge, and the s1 experiment showed what a 1,000-example fine-tune plus extra thinking time could do.
Long-term threads
Europe’s program continued with the AI Continent Action Plan in April, the Code of Practice for general-purpose models in July and the obligations for those models in August. OpenAI’s chip work resurfaced in September reports of a chip built with Broadcom and in October’s 10-gigawatt agreement with the company. The Musk dispute moved to court with OpenAI’s countersuit in April. The Cohere case was an early sign of the copyright disputes that ran through the year, from June’s fair-use rulings on generative-AI training to Anthropic’s settlement in September. And DeepSeek’s trouble with privacy regulators, begun in Italy in January, now reached South Korea.