PAUL MCQUESTEN paulmcq

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.

Stevey's Google Platforms Rant

I was at Amazon for about six and a half years, and now I've been at Google for that long. One thing that struck me immediately about the two companies -- an impression that has been reinforced almost daily -- is that Amazon does everything wrong, and Google does everything right. Sure, it's a sweeping generalization, but a surprisingly accurate one. It's pretty crazy. There are probably a hundred or even two hundred different ways you can compare the two companies, and Google is superior in all but three of them, if I recall correctly. I actually did a spreadsheet at one point but Legal wouldn't let me show it to anyone, even though recruiting loved it.

I mean, just to give you a very brief taste: Amazon's recruiting process is fundamentally flawed by having teams hire for themselves, so their hiring bar is incredibly inconsistent across teams, despite various efforts they've made to level it out. And their operations are a mess; they don't real

	"""
	The most atomic way to train and run inference for a GPT in pure, dependency-free Python.
	This file is the complete algorithm.
	Everything else is just efficiency.

	@karpathy
	"""

	import os # os.path.exists
	import math # math.log, math.exp

	import asyncio
	import json
	import time

	from typing import Optional, List

	from pydantic import BaseModel, Field

	from starlette.responses import StreamingResponse
	from fastapi import FastAPI, HTTPException, Request

	//
	// Created by farhang on 21.01.24.
	//


	#include <iostream>
	#include <coroutine>
	#include <exception>
	#include <cstdlib>

	// Daniel Lemire in this article
	// https://lemire.me/blog/2013/12/26/fastest-way-to-compute-the-greatest-common-divisor/
	// presented benchmark comparisons of different implementations of Stein's (binary) gcd algorithm.
	// The wikipedia (iterative) implementation of the algorithm was noted to be very inefficient.
	// Lemire presented benchmarked comparisons of various implementations, original code below.
	// https://github.com/lemire/Code-used-on-Daniel-Lemire-s-blog/blob/master/2013/12/26/gcd.cpp
	// I have modified the output to make it explicit and clear, modified the code in some functions
	// to make them easier to follow and standardized variable names, and added the Ruby implementation.
	// I also ordered the ouput to show fastest to slowest.
	// The results: gcdwikipedia7fast32 is fastest by far; the implementation shown in widipedia is slowest.