SUMM

Speaker jokes with audience to warm up and identifies Replicate employees in the room
Zeke Sikelianos introduces himself, shares his GitHub/X handle, and describes Replicate as a cloud platform for running AI models from open source and proprietary sources, including custom ones

Karpathy is an AI researcher with experience at Google, OpenAI, Tesla, and more recently Eureka Labs, an educational platform
He is notable for his accessible YouTube talks, popularizing concepts like "vibe coding" and describing "English" as the hottest programming language
Karpathy wrote the "software 2.0 manifesto," predicting a future where machine learning models write code more effectively than humans

Karpathy created MenuGen (Menuguen), a web app for generating menu item images from text, useful for those unfamiliar with menu languages or terminology
Development was fun locally but became much more complex and frustrating when deployed due to multiple platform pain points
Karpathy published a blog post describing his frustrations, highlighting issues with outdated documentation, API changes, rate limits, and onboarding—even for paid users
Replicate was acknowledged alongside major companies, underscoring shared developer experience shortcomings

Replicate teams embraced "llm.text" to provide documentation in markdown and plaintext formats, making it easy for LLMs to parse
Added features: model pages have a button to copy markdown docs or send them directly to Claude or ChatGPT for interactive exploration
The approach makes documentation more interactive for both humans and LLMs

Karpathy emphasizes that LLMs prefer API documentation as simple curl commands rather than complex interfaces
Curl, despite being old and basic, is accessible and fully describes an API request, making it ideal for both humans and LLMs

Cog, an open-source tool, helps package machine learning models into Docker containers with standardized APIs
All Cog docs are available in a single llm.ext file, so editors can consume them for context—enabling LLMs to use and modify code more effectively

The primary audience for product documentation and APIs is now language models, not just human developers
Having machine-readable, up-to-date OpenAPI schemas is essential to ensure LLMs can access and understand API functionalities

MCP (Model-Connected Plugin, inferred from context) bridges OpenAPI schemas and LLMs, allowing LLMs to interact directly with APIs
Installation is straightforward, involving only an API token and a line of JSON in dev settings; enables tools like Claude, GitHub Copilot, and VS Code to use Replicate's API through MCP
MCP facilitates discovery, scaffolding, and interactive development via language models, powered by a well-documented OpenAPI schema

Accept payments flexibly: blocking legitimate new users for rapid API use is a mistake—enable credit-based or pay-to-unlock systems
Always document new features before considering them done; documentation now also targets LLM consumption
Output documentation and data in formats most accessible to LLMs (markdown/plaintext rather than flashy HTML)
Use established, "boring" technologies so LLMs trained on conventional software can generate/use code effectively
Design APIs with concise, information-dense outputs that are easy for LLMs to handle (avoid overloading with unnecessary data)

For generating docs: start with well-defined OpenAPI schemas in YAML or JSON; plenty of tools (e.g., Docusaurus, ReadTheDocs, ReadMe.com) transform schemas into end-user and SDK docs
Discovery and purchasing: ensure APIs provide access to key decision data (e.g., model pricing), enabling LLMs to compare and recommend products/services based on structured API information

Design like Karpathy is watching - Zeke Sikelianos, Replicate