OPEN AI'S MIDTRAINING ARCHITECT REBUILDS THE BRAIN

Andrej Karpathy returns to OpenAI to teach models new tricks without starting from scratch

Andrej Karpathy, the Slovak-born AI pioneer whose work powers Tesla’s Autopilot and OpenAI’s GPT lineage, has quietly returned to OpenAI to lead a small team focused on midtraining—teaching base models new skills, behaviors, and factuality without full retraining. The move signals a pivot from frontier model scaling to the operational art of making existing models smarter, safer, and more useful in real deployments. 'Pretraining is mostly solved at the frontier,' Karpathy told *The Neural Times*. 'The next leap is midtraining: synthetic data, targeted fine-tuning, and alignment without the compute burn of a full retrain.'