What topics does this roadmap cover?

Karpathy's AutoResearch: Foundations, The Karpathy Loop: How the Agent Runs Experiments, Writing Great program.md, Scaling AutoResearch: SkyPilot, AgentHub, and Consumer GPUs, Real Results, Failure Modes, and Ecosystem

How many units are in this roadmap?

This roadmap contains 106 units across 5 sections.

Is this roadmap free?

Yes, this roadmap is completely free.

Karpathy's AutoResearch - Roadmaps

##### ###### ##### ### # # ### # # ###### ## ## ## ## ## ## ## # # # # # ## ##### #### ##### # # # # # # # #### ## # ## ## ## ## # # # # # ## ## # ###### ## ### # ### # ######

Writing Great program.md

How to write research instructions that produce useful overnight runs instead of wasted compute.

3.1Introduction

What's ahead

0/1 tasks2 minutes

3.2Why program.md Is Your Main Tool Now

English, not Python

0/1 tasks4 minutes

3.3The Five Sections of program.md

Five-part structure

0/1 tasks4 minutes

3.4The Objective Statement

One metric, one direction

0/1 tasks3 minutes

3.5Defining What Can Be Modified

The file boundary

0/1 tasks3 minutes

3.6Defining What Cannot Be Touched

Three things off-limits

0/1 tasks3 minutes

3.7The Simplicity Criterion

Complexity vs. improvement

0/1 tasks4 minutes

3.8Setting VRAM Constraints

Soft memory constraint

0/1 tasks4 minutes

3.9The Evaluation Directive

Score extraction logic

0/1 tasks3 minutes

3.10Scope Too Narrow

Grid search in disguise

0/1 tasks4 minutes

3.11Scope Too Wide

Wasted experiments

0/1 tasks4 minutes

3.12Calibrating Time Budget to Hardware

Hardware-specific tuning

0/1 tasks4 minutes

3.13Directing the Agent Toward Hypotheses

Focused starting points

0/1 tasks4 minutes

3.14What to Tell the Agent About Prior Results

Saving agent time

0/1 tasks4 minutes

3.15program.md for Non-ML Problems

Beyond ML training

0/1 tasks4 minutes

3.16Case Study: Karpathy's Original program.md

Minimal and effective

0/1 tasks4 minutes

3.17Case Study: Shopify's Adaptation

Ruby, not Python

0/1 tasks4 minutes

3.18What Separates a Good Run from a Wasted One

Diverse vs. repetitive

0/1 tasks4 minutes

3.19Iterating on program.md Between Runs

Refining between runs

0/1 tasks4 minutes

3.20Common Mistakes

Five frequent errors

0/1 tasks4 minutes

3.21Section Recap

What you've learned

0/1 tasks2 minutes

3.1Introduction

What's ahead

0/1 tasks2 minutes

3.2Why program.md Is Your Main Tool Now

English, not Python

0/1 tasks4 minutes

3.3The Five Sections of program.md

Five-part structure

0/1 tasks4 minutes

3.4The Objective Statement

One metric, one direction

0/1 tasks3 minutes

3.5Defining What Can Be Modified

The file boundary

0/1 tasks3 minutes

3.6Defining What Cannot Be Touched

Three things off-limits

0/1 tasks3 minutes

3.7The Simplicity Criterion

Complexity vs. improvement

0/1 tasks4 minutes

3.8Setting VRAM Constraints

Soft memory constraint

0/1 tasks4 minutes

3.9The Evaluation Directive

Score extraction logic

0/1 tasks3 minutes

3.10Scope Too Narrow

Grid search in disguise

0/1 tasks4 minutes

3.11Scope Too Wide

Wasted experiments

0/1 tasks4 minutes

3.12Calibrating Time Budget to Hardware

Hardware-specific tuning

0/1 tasks4 minutes

3.13Directing the Agent Toward Hypotheses

Focused starting points

0/1 tasks4 minutes

3.14What to Tell the Agent About Prior Results

Saving agent time

0/1 tasks4 minutes

3.15program.md for Non-ML Problems

Beyond ML training

0/1 tasks4 minutes

3.16Case Study: Karpathy's Original program.md

Minimal and effective

0/1 tasks4 minutes

3.17Case Study: Shopify's Adaptation

Ruby, not Python

0/1 tasks4 minutes

3.18What Separates a Good Run from a Wasted One

Diverse vs. repetitive

0/1 tasks4 minutes

3.19Iterating on program.md Between Runs

Refining between runs

0/1 tasks4 minutes

3.20Common Mistakes

Five frequent errors

0/1 tasks4 minutes

3.21Section Recap

What you've learned

0/1 tasks2 minutes

Save Progress

#####   ######  #####    ###    #   #  ###  #   #  ######
##  ##  ##      ##  ##  ## ##   #   #   #   #   #  ##
#####   ####    #####   #   #   #   #   #   #   #  ####
##  #   ##      ##      ## ##    # #    #    # #   ##
##   #  ######  ##       ###      #    ###    #    ######

$ curl repovive.com/roadmaps/karpathy-autoresearch

░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░██████████████████████████████████████████████████

Browse Karpathy's AutoResearch Topics

Explore individual topics with detailed lessons and practice problems.

Karpathy's AutoResearch: Foundations

Install AutoResearch, understand the three-file architecture, and run your first autonomous overnight experiment loop.

22 lessons78 min

The Karpathy Loop: How the Agent Runs Experiments

How the agent reasons, edits code, runs training, and decides whether to keep or revert each experiment.

24 lessons