Modal demonstrates how AI research agents can leverage elastic GPU provisioning to run experiments efficiently. Using Karpathy's autoresearch framework and Claude Code equipped with Modal Skills, an agent tackled OpenAI's Parameter Golf challenge autonomously over 15 hours — running 113 experiments across 238 GPU-hours. The agent dynamically scaled from single-GPU sandboxes for debugging to 40 simultaneous H100s for validation, achieving 5x speedup over a single workstation while using a fraction of a dedicated cluster's resources. Modal's serverless runtime lets agents provision GPUs via simple API calls, scaling to zero when idle to avoid wasted spend.
Sort: