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.

DMontgomery40

codex-local-rag

codex-local-rag builds a local hybrid search index over Codex history so you can ask things like:

• "Where did I debug that nginx reverse proxy issue?"
• "How did I fix a failing GitHub Actions workflow last month?"
• "What command did I use when I recovered that SQLite database?"

It is designed for real Codex artifacts, not generic chat logs. Instead of slicing everything into arbitrary chunks, it reconstructs turns, keeps tool output separate when it gets large, and combines sparse plus dense retrieval for better recall on both conceptual questions and exact technical strings.

rexim

#include <stdio.h>
#include <stdlib.h>

#define da_append(xs, x)                                                             \
    do {                                                                             \
        if ((xs)->count >= (xs)->capacity) {                                         \
            if ((xs)->capacity == 0) (xs)->capacity = 256;                           \
            else (xs)->capacity *= 2;                                                \
            (xs)->items = realloc((xs)->items, (xs)->capacity*sizeof(*(xs)->items)); \
        }                                                                            \

afiodorov

Docker Image  : pytorch/pytorch
Image Runtype : jupyter_direc ssh_direc ssh_proxy
Environment : [["JUPYTER_DIR", "/"], ["-p 41654:41654", "1"]]

pip install torch bitsandbytes sentencepiece "protobuf<=3.20.2" git+https://github.com/huggingface/transformers flask python-dotenv Flask-HTTPAuth accelerate

!mv /opt/conda/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cuda116.so /opt/conda/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cpu.so

saint11

-- BASIC EASING FUNCIONS CHEAT CHEET
-- (most of them are based on glide by Jacob Albano https://github.com/jacobalbano/glide)

function ease_linear(t)
	return t
end

function ease_cube_in(t)
	return t * t * t
end

skeeto

/* Avalanche matrix visualizer
 * $ cc -Ofast -fopenmp -Wall -Wextra avalanche.c
 * This is free and unencumbered software released into the public domain.
 */
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>

#define NSAMPLES (1L << 24)
#define SCALE    24

YukiSnowy

// Windows
// g++ *.cpp -o vulkan -lSDL2main -lSDL2 -lvulkan-1
// Linux
// g++ *.cpp -o vulkan -lSDL2main -lSDL2 -lvulkan
// https://vulkan-tutorial.com/

#include <iostream>
using namespace std;

#include <SDL2/SDL.h>

llandsmeer

#include <stdio.h>
#include <stdlib.h>
#include <libtcc.h>

void callme(int x) {
    printf("hello, %d\n", x);
}

void error_func(void * user, const char * msg) {
    printf("TCC Error: %s\n", msg);

SemanticDevice

#include <allegro5/allegro5.h>
#include <allegro5/allegro_primitives.h>
#include <stdio.h>
#include <stdlib.h>

#define WIN_WIDTH_PX (800)
#define WIN_HEIGHT_PX (800)
#define FPS (60.0f)

static void Initialize();

munificent

#include <stdio.h>
#include <stdlib.h>

const int H = 40;
const int W = 80;

char map[H][W];

int rnd(int max) {
  return rand() % max;