Google Home's smart assistant just got a significant brain transplant. Gemini 3.1 now understands and executes multi-step commands, a leap for conversational AI in the home.
A new approach to Retrieval Augmented Generation (RAG) is making waves. It ditches vector embeddings entirely, achieving near-perfect results on a key benchmark.
Forget the latest LLM buzzwords. This deep dive goes back to the absolute basics: character-based tokenization. Turns out, the old ways might be smarter than we think.
Large language models are exhibiting multilingual mixing, a phenomenon dubbed 'code-switching.' This isn't random noise; it's a complex behavior with deep roots in their training data and architecture.
Users expect conversational fluidity, not just spoken text. Amazon's Nova 2 Sonic aims to bridge the gap between static text agents and dynamic voice assistants, but the market's readiness is a complex question.
Think AI understands everything? Think again. A deep dive into the hidden linguistic battleground where Italian's unique grammar was tripping up even the smartest models.