AI digest: Hardware wars heat up
Mistral launches unified multimodal model, Nvidia pushes trillion-dollar chip projections, and legal battles mount over AI training data.
Big week for model releases and hardware announcements, plus the usual copyright drama heating up.
Mistral Small 4 consolidates everything into one model
Mistral AI dropped Mistral Small 4, a 119B-parameter mixture of experts model that handles instruction following, reasoning, and multimodal tasks in one deployment. Smart move to reduce the operational complexity of running multiple specialised models. Makes you wonder if this unified approach will become the standard rather than having separate models for different tasks.
Jensen Huang claims $1 trillion in chip orders
Nvidia’s CEO said he expects $1 trillion worth of orders for Blackwell and Vera Rubin chips. That’s an absolutely bonkers number that suggests either unprecedented AI infrastructure buildout or classic Silicon Valley hyperbole. Meta’s $27 billion deal with Nebius for some of those Vera Rubin chips shows the spending is real though.
Copyright lawsuits pile up for OpenAI
Encyclopedia Britannica and Merriam-Webster are suing OpenAI over nearly 100,000 allegedly copied articles used for training. This comes alongside a child safety lawsuit against xAI over Grok’s image generation capabilities. The legal pressure is mounting from multiple angles, which could reshape how these models get trained and deployed.
GPT-4.5 passes Turing test by playing dumb
Researchers got GPT-4.5 to fool 73% of people into thinking it was human by telling it to make typos and act less capable. Says something interesting about what makes us seem human, and raises questions about future AI detection methods.