Design like Karpathy is watching

Replicate · Beginner ·🧠 Large Language Models ·1y ago

Skills: API Design90%LLM Engineering80%

Key Takeaways

This video teaches API design principles and best practices for large language models

Original Description

Update, July 22 2025: The official version of this talk is now live at https://www.youtube.com/watch?v=huQPkrwVWwc This video delves into the evolving landscape of API design, emphasizing the need to cater to large language models (LLMs) as primary consumers, beyond human developers. It highlights the challenges faced when AI models interact with traditional APIs, drawing insights from Andrej Karpathy's "vibe coding" experience with MenuGen. Learn about practical strategies like embracing `llms.txt` for machine-readable documentation, leveraging established "boring" technologies (like SQL, Python, and React), and practicing good API hygiene. GitHub repo: https://github.com/zeke/design-like-karpathy-is-watching 00:00 Intro and Replicate overview 01:00 Who is Andrej Karpathy? 02:11 What is MenuGen? 02:49 The pains of vibe coding deployed apps 03:17 Replicate's API challenges 04:31 Embrace llms.txt 06:18 Why LLMs like to curl 07:46 Cog's llms.txt and AI code generation 08:44 LLMs as primary audience 09:18 What is MCP (Model Context Protocol)? 10:15 Replicate's MCP server 11:02 Claude Desktop with Replicate MCP 12:36 Lessons learned for API design 15:56 Conclusion

Watch on YouTube ↗ (saves to browser)

Sign in to unlock AI tutor explanation · ⚡30

More on: API Design

View skill →

Go API Tutorial - Make An API With Go

Go API Tutorial - Make An API With Go

Build Login/Register API Server w/ Authentication | JWT Express AUTH using Passport.JS and Sequelize

Build Login/Register API Server w/ Authentication | JWT Express AUTH using Passport.JS and Sequelize

GraphQL + WordPress (P4D3) - Live Coding with Jesse

GraphQL + WordPress (P4D3) - Live Coding with Jesse

freeCodeCamp.org

Python Flask Tutorial: Full-Featured Web App Part 10 - Email and Password Reset

Python Flask Tutorial: Full-Featured Web App Part 10 - Email and Password Reset

Full Socket.io and React.js Online Multiplayer Tic-Tac-Toe Game | Socket.io From Zero To Hero

Full Socket.io and React.js Online Multiplayer Tic-Tac-Toe Game | Socket.io From Zero To Hero

Spring Boot Project: Build a REST API for an E-commerce Platform

Spring Boot Project: Build a REST API for an E-commerce Platform

Programming with Mosh

Related Reads

Introducing Claude Opus 5 on AWS: Anthropic’s most capable Opus model

Learn about Claude Opus 5, Anthropic's most capable Opus model, and how to integrate it into agentic systems on AWS

AWS Machine Learning

Can AI Keep a Great Mind Alive?

Learn how AI can preserve a great mind's thinking through persona fine-tuning, first-principles reasoning, and mechanistic interpretability

Anthropic launches Opus 5

Anthropic's Opus 5 offers a cheaper and less restrictive alternative to Fable, making it a preferable choice for most use cases

Claude Opus 5 arrives with near Fable performance at half the price

Learn about Claude Opus 5's upgraded features and improved performance at a lower price point, making it an attractive option for developers and enterprises

Chapters (14)

Intro and Replicate overview

1:00 Who is Andrej Karpathy?

2:11 What is MenuGen?

2:49 The pains of vibe coding deployed apps

3:17 Replicate's API challenges

4:31 Embrace llms.txt

6:18 Why LLMs like to curl

7:46 Cog's llms.txt and AI code generation

8:44 LLMs as primary audience

9:18 What is MCP (Model Context Protocol)?

10:15 Replicate's MCP server

11:02 Claude Desktop with Replicate MCP

12:36 Lessons learned for API design

15:56 Conclusion

5 Levels of AI Agents - From Simple LLM Calls to Multi-Agent Systems

Dave Ebbelaar (LLM Eng)