Keywords：LLM cognitive core, Edge AI, Small-scale models, Personal computing, Multimodal, On-device fine-tuning, Cloud-edge collaboration, Gemma 3n model, Billion-parameter models, AI capability adjustment, Tool utilization capability, Always-on AI
🔥 Focus
Karpathy proposes the “LLM cognitive core” concept, on-device AI competition intensifies: Andrej Karpathy predicts that the next focus of the AI race will be the “LLM cognitive core”—a small model with billions of parameters that sacrifices encyclopedic knowledge for ultimate capabilities. It will serve as the kernel of personal computing, always-on and running by default on every device. Its features include being natively multimodal, having adjustable capabilities, powerful tool use, on-device fine-tuning, and collaboration with large cloud-based models. This vision coincides with Google’s release of the Gemma 3n model, which supports multimodal input and requires only 2GB of RAM to run, demonstrating the potential for powerful AI capabilities on-device and signaling that the era of personal AI computing is accelerating. (Source: karpathy, Omar Sanseviero, zacharynado, teortaxesTex, jeremyphoward)