Anthropic Tasks Karpathy With Recursive AI Pre-Training

Former OpenAI co-founder Andrej Karpathy has joined Anthropic’s pre-training team to focus on recursive self-improvement. Reporting to Head of Pre-training Nick Joseph, Karpathy is tasked with building a dedicated internal group that uses Claude to automate and optimize foundational AI research. The objective is to use the current generation of models to accelerate the compute-intensive training runs of their successors.

The Agentic Engineering Context

This appointment operationalizes Karpathy’s vision for “Software 3.0”, an architecture where the large language model acts as the primary host process and delegates tasks to traditional software co-processors. In early 2025, Karpathy popularized vibe coding, a paradigm where developers specify intent in natural language while AI agents handle the underlying logic generation.

By January 2026, this concept evolved into structured agentic engineering workflows. Karpathy heavily promoted the CLAUDE.md configuration standard for managing agent behavior across complex repositories. This standard surpassed 220,000 stars on GitHub by May 2026, signaling broad developer acceptance of repository-level agent orchestration. His new role at Anthropic will likely scale these exact workflows to the massive codebases required for frontier model training.

Automating Pre-Training Infrastructure

Pre-training a frontier model requires orchestrating tens of thousands of GPUs, managing petabytes of data, and constantly adjusting hyperparameters. Traditionally, this requires large teams of specialized infrastructure engineers. The new team’s mandate involves using existing Claude models to write the necessary training scaffolding, optimize distributed communication algorithms, and evaluate data quality pipelines automatically.

To dedicate his time to this effort, Karpathy has paused operations at Eureka Labs, the AI-native education startup he founded in July 2024. His return to frontier model development brings significant systems engineering experience. He previously spent five years directing Tesla’s Autopilot computer vision division and spent 2023 to 2024 at OpenAI developing mid-training and synthetic data pipelines.

Broader Platform Consolidation

Anthropic is actively securing the supporting infrastructure required for these autonomous research workflows. The day before announcing Karpathy’s hire, the company acquired Stainless for $300 million. This acquisition gives Anthropic in-house control over automated SDK generation and critical server connectors for the Model Context Protocol. To balance these new capabilities with safety, the company concurrently hired former Meta engineer Chris Rohlf to lead frontier red teaming and stress-test these highly autonomous systems.

If you build agentic workflows, the convergence of automated pre-training and native SDK generation dictates your integration strategy. As AI models successfully automate their own backend infrastructure, the highest value work will shift entirely to context engineering and system orchestration.