Skip to content

01-the-idea.md

Latest commit

 

History

History
47 lines (38 loc) · 2.53 KB

01-the-idea.md

File metadata and controls

47 lines (38 loc) · 2.53 KB
Raw

Chapter 1: The Idea — "Can We Run Autoresearch on AWS for Pennies?"

Time: 15 minutes Cost: $0.00 Key Insight: Start with a clear goal and let the AI interview you to refine it.

Context

Karpathy released autoresearch — an autonomous ML research framework where AI agents modify training code, run 5-minute experiments, and iterate overnight. But it assumes you have an H100 GPU. We wanted to run it on AWS without owning any GPU.

The Prompt

https://github.com/karpathy/autoresearch 를 사용해서 딥러닝 모델 성능 개선 실험을
재현해 보고 싶어. OMC의 autoresearch를 사용해서 실험을 할 수 있게 셋팅해줘.
추가 컨텍스트가 필요하면 심층 인터뷰를 해줘.

Translation: "I want to reproduce deep learning model performance improvement experiments using https://github.com/karpathy/autoresearch. Set it up so I can run experiments with OMC's autoresearch. If you need additional context, do a deep interview."

What Happened

The AI started a deep interview to understand the requirements:

  1. Computing environment? → SageMaker Managed Spot Training (serverless, pay per use)
  2. Agent loop? → OMC autopilot mode (autonomous)
  3. Goal? → Reproduce experiments first, then create education/demo materials

This interview shaped the entire project direction. Instead of just setting up autoresearch, we designed a parallel evolution pipeline — running multiple experiments simultaneously on cheap Spot instances.

The Result

A clear project plan emerged:

  • Parallel evolution pipeline (not sequential like the original)
  • SageMaker Spot Training for cost efficiency
  • HUGI pattern (Hurry Up and Get Idle) for zero idle cost
  • Every experiment documented as a tutorial

Lessons Learned

  • Always ask for a deep interview — the AI discovered that SageMaker Spot was the right approach, not EC2 instances or local GPUs
  • "Reproduce" became "improve" — the interview revealed the real goal was education + cost optimization, not just reproduction
  • The prompt that changes everything: Adding "추가 컨텍스트가 필요하면 심층 인터뷰를 해줘" ("Do a deep interview if you need more context") turned a vague request into a focused plan

Try It Yourself

# Start Claude Code in any project and try:
# "I want to [your goal]. Do a deep interview to refine the approach."

Running Cost

Phase Action Cost Cumulative
Planning Deep interview + architecture design $0.00 $0.00