karpathy

LLM Wiki

A pattern for building personal knowledge bases using LLMs.

This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.

The core idea

Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.

fasterthanlime

// Update 2025-05-22 — ported to rand v0.9, using a struct with Display impl rather than returning a string

struct MarineLine;

impl std::fmt::Display for MarineLine {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        use rand::{Rng, seq::IndexedRandom};

        let marine_life = ["🐳", "🐠", "🦈", "🐙", "🐡", "🐬", "🐟", "🦀", "🐋"];
        let water = [

w00kie

import { z } from 'zod';
import { ZipWriter, BlobReader, configure } from '@zip.js/zip.js';

// Without this, we get uncaught error due to Workers runtime bug
// See: https://github.com/gildas-lormeau/zip.js/discussions/514
configure({
	useCompressionStream: false,
});

// Payload schema that lists the files to be bundled, their filenames and the archive filename

mattfysh

import fs from 'node:fs/promises'
import path from 'node:path'
import rimraf from 'rimraf'
import { glob } from 'glob'
import { generate } from 'ts-to-zod'
import ts from 'typescript'
import { v4 as uuidv4 } from 'uuid'

const OUTDIR = 'packages/types/gen'
const { factory } = ts

crohr

# .github/workflows/pullpreview.yml
name: PullPreview
on:
  # the schedule is optional, but helps to make sure no dangling resources are left when GitHub Action does not behave properly
  schedule:
    - cron: "30 2 * * *"
  # optional, only use if you want to have an always-on branch
  push:
    branches:
      - main

yonatanzunger

import dis
import io
from types import CodeType, FunctionType
from typing import Any, Callable, List, NamedTuple, Tuple

# Opcodes
_LOAD_FAST = dis.opname.index('LOAD_FAST')
_LOAD_CONST = dis.opname.index('LOAD_CONST')
_COMPARE_OP = dis.opname.index('COMPARE_OP')
_JUMP_IF_FALSE_OR_POP = dis.opname.index('JUMP_IF_FALSE_OR_POP')

mmazzarolo

/**
 * RuntimeGlobalsChecker
 * 
 * You can use this utility to quickly check what variables have been added (or
 * leaked) to the global window object at runtime (by JavaScript code). 
 * By running this code, the globals checker itself is attached as a singleton
 * to the window object as "__runtimeGlobalsChecker__". 
 * You can check the runtime globals programmatically at any time by invoking
 * "window.__runtimeGlobalsChecker__.getRuntimeGlobals()".
 * 

erikw

git commit-status alias

An alias that will generate a git commit message staged changes as shown in git-status. Put this alias (section below) in your .gitconfig.

The message generated will be in the format of:

$ git status --porcelain
A file1.py
A file2.py
A file3.py

roustem

import os
import re

date_re = re.compile("(\d\d\d\d)(\d\d)(\d\d).md")

for root, dirs, files in os.walk("./"):
    for filename in files:
        if filename.endswith(".md"):
            new_filename = filename
            m = date_re.match(filename)

nuintun

/**
 * @module compose
 */

export interface Next {
  (): Promise<void>;
}

export interface Composed<C> {
  (context: C, next?: Next): Promise<void>;