Karpathy is at it again. This time, it's not a paper or a gist—he's built a social network directly.

[KarpathyTalk](https://karpathytalk.com/) was recently open-sourced, with an interesting positioning: a combination of X and GitHub Gists. Its official definition is "an open social network for the LLM era."


## What It Solves

The default stance of traditional social platforms is: lock content and relationships in databases, make APIs expensive and hard to use, and require crawlers to take detours.

KarpathyTalk is the complete opposite. It assumes your content will be freely accessed by crawlers, Agents, and third-party tools. All data is directly exposed via API in two formats: JSON for code and LLMs to parse structurally, and Markdown for humans to read directly. No API key barriers, no rate limits, no data locking.

This is the fundamental difference between it and ordinary social products.

## How It Works

Log in directly with GitHub OAuth—no need to remember a new password. The editor is Gist-style, natively supports Markdown, automatically syntax-highlights code blocks, and allows direct image drag-and-drop. Each post is limited to 10,000 characters—not a lot, but enough to express a complete opinion while forcing conciseness.

The social layer is restrained: likes, reposts, quote replies, and follows. Features are sufficient to support the community but not bloated.

The tech stack is also simple: Go single binary + SQLite + htmx + goldmark. It takes just a few commands to run locally.


## Notable Details

The platform has a statement: Posts are by default from humans. AI-generated or AI-assisted content is allowed, but must be tagged with #AI.

Additionally, Karpathy himself wrote a sentence on the project page: "I can't 100% guarantee the long-term survival of this website. If you invest content on it, you'd better back it up yourself." — This is an experiment, so you should have expectations before using it.

## What to Think About

There have been many social products in the past two years, but most are competing on the functional level. KarpathyTalk has a different approach—it doesn't compete on features, but takes a stance on the data model. Whether open data and lowering the access threshold for Agents have practical value is uncertain, but it does offer a different possibility.

github.com/karpathy/KarpathyTalk