Karpathy leaves OpenAI and joins Anthropic to train Claude

Andrej Karpathy, renowned artificial intelligence researcher and co-founder of OpenAI, has switched sides. He is now joining Anthropic to focus on the pre-training of Claude, its flagship model. This phase is the most resource-intensive and defines the system's core capabilities, marking a key move in the competition between AI labs.

technical illustration of neural network pre-training process, Andrej Karpathy figure standing beside a massive glowing GPU server rack running Claude model training, cascading data streams flowing into multi-layered transformer architecture, gradient descent pathways visualized as luminous blue and orange lines converging, server cooling pipes and fiber optic cables connected to high-density compute nodes, cinematic engineering visualization, dark server room atmosphere with cold blue ambient light, motion blur on rotating hard drives, real-time loss curve display on holographic monitor, photorealistic industrial render, ultra-detailed hardware components

Pre-training with Claude: the phase that defines the knowledge base 🧠

Karpathy will form a team to optimize this critical stage, where the model absorbs massive data to acquire its foundations. The irony is that he will use Claude itself as an acceleration tool, creating a feedback loop. This approach aims to reduce costs and improve efficiency, but also raises questions about biases and control in the generation of synthetic knowledge during the process.

The scientist who set out to train his own boss 🤖

So Karpathy, after helping to birth GPT at OpenAI, is now dedicated to making Claude smarter. And the best part: he will use Claude so that Claude learns faster. It's like a professor asking his most diligent student to help him prepare the next day's class. Let's hope Claude doesn't decide to charge him overtime for consulting.