I trained gpt-1 on my local machine (RTX 2060 Super 8GB VRAM)

In three linesUser trained GPT-1 on RTX 2060 Super (8 GB VRAM) in ~1 hour using Claude-generated code based on original implementation. Cost to reproduce GPT models dropped 500–1000× since GPT-2 ($43,000 → $48 per H100 cluster run).

Summary generated by Claude — human-verified