Karpathy defines the LLM 'Cognitive Core': On-device, tool-using, and trainable

cephaloform · x · 2026-08-18

Andrej Karpathy outlines the vision for an LLM 'Cognitive Core': a few-billion-parameter model that sacrifices encyclopedic knowledge for capability, running locally as a kernel for personal computing. Key features include native multimodality, Matryoshka-style scaling for capability, reasoning, aggressive tool use, and on-device LoRA slots for personalization. It delegates complex checks to cloud oracles when online.

Original post →

More from Models

Models channel →