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.

antenore

Configure OpenAI Codex CLI with DeepSeek Models

Disclaimer (May 2026). Sorry for the long radio silence on the comments, busy stretch on my side. I no longer use this setup and stopped tracking it actively. I'm refreshing the page out of thoroughness because comments kept coming in and people deserve an honest snapshot. If you find better workarounds, post them in the comments and I'll link them, but I won't be maintaining this further.

Status

Codex deprecated wire_api = "chat" in favour of OpenAI's Responses API. DeepSeek only speaks Chat Completions, so the direct integration documented below is no longer reliable. Symptoms reported by users:

• An assistant message with 'tool_calls' must be followed by tool messages... (insufficient tool messages following tool_calls)
• Missing 'reasoning_content' field in the assistant message at message index N

padeoe

🤗Huggingface Model Downloader

Note

(2026-06-18) Add 📊Download Progress, 🔁integrity-aware resume (re-fetch missing or changed files), and fast, resumable listing for 📚repos with massive file counts.
(2025-01-08) Add feature for 🏷️Tag(Revision) Selection, contributed by @Bamboo-D.
(2024-12-17) Add feature for ⚡Quick Startup and ⏭️Fast Resume, enabling skipping of downloaded files, while removing the git clone dependency to accelerate file list retrieval.

Considering the lack of multi-threaded download support in the official huggingface-cli, and the inadequate error handling in hf_transfer, This command-line tool leverages curl and aria2c for fast and robust downloading of models and datasets.

Features

geekodour

This is comparison between whisper.cpp and faster-whisper. The faster-whisper readme has some benchmarks on the readme but wanted to test it myself. For whisper, I just ran manually. For faster-whisper, wrote this small script

• https://github.com/guillaumekln/faster-whisper/
• https://github.com/ggerganov/whisper.cpp
• quanitization w faster-whisper: https://opennmt.net/CTranslate2/quantization.html
• quantization w whisper.cpp: https://github.com/ggerganov/whisper.cpp#quantization

whisper.cpp

• ./main -bs 5 -p 2 -f steve2.wav -m models/ggml-small.en.bin
  • Total 8 CPU threads on my 12 core machine
  • -bs 2 : actually performs better about 10s faster.

garg-aayush

Steps to manage multiple CUDA environments

Latest Update: May 19th, 2024

This gist contains all the steps required to:

• Install multiple CUDA versions (e.g., CUDA 11.8 and CUDA 12.1
• Manage multiple CUDA environments on Ubuntu using the utility called environment modules.
• Use this approach to avoid CUDA environment conflicts.

Environment Modules is a package that provides for the dynamic modification of a user's environment via modulefiles. You can find more on it at https://modules.readthedocs.io/en/latest/

mattiasarro

# Taken from https://johanwind.github.io/2023/03/23/rwkv_details.html. 
# I've added additional comments restructured it a tiny bit, which makes it clearer for me.

import numpy as np
from torch import load as torch_load  # Only for loading the model weights
from tokenizers import Tokenizer

exp = np.exp
layer_norm = lambda x, w, b : (x - np.mean(x)) / np.std(x) * w + b
sigmoid = lambda x : 1/(1 + exp(-x))

ymoslem

# This example uses M2M-100 models converted to the CTranslate2 format.
# Download CTranslate2 models:
# • M2M-100 418M-parameter model: https://bit.ly/33fM1AO
# • M2M-100 1.2B-parameter model: https://bit.ly/3GYiaed


import ctranslate2
import sentencepiece as spm

tommyip

# Based on https://gist.github.com/bastibe/c0950e463ffdfdfada7adf149ae77c6f
# Changes:
# * Instead of overriding cd, we detect directory change. This allows the script to work
#   for other means of cd, such as z.
# * Update syntax to work with new versions of fish.
# * Handle virtualenvs that are not located in the root of a git directory.

function __auto_source_venv --on-variable PWD --description "Activate/Deactivate virtualenv on directory change"
  status --is-command-substitution; and return

SteveHere

import logging
import socks # use pysocks
import asyncio
from datetime import datetime
from itertools import cycle

logging.basicConfig(level=logging.INFO)
socks_router_loop = cycle((  # simple round-robin router to socks proxies
    ('127.2.0.0', 9050, None, None),  # address, port, username, password
    # ('127.3.0.0', 9050, "proxy", "passwordpassword"),

cristianadam

# MIT License
# 
# Copyright (c) 2019 Cristian Adam
# 
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions: