Adam McNight McNight

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.

QEMU on Apple Silicon with Vulkan support

The following is taken from https://gist.github.com/akihikodaki/87df4149e7ca87f18dc56807ec5a1bc5 with modifications to introduce Vulkan support. Since we need custom patches to QEMU, virglrenderer, MoltenVK, and libepoxy, the provided script will pull the right refs and build everything with the right params. Feel free to inspect the script to find the source for all the modifications. We are actively working to upstream everything so please do not build anything long-term with these patches.

Testing

Prerequisites

Xcode (relatively recent version)
Vulkan SDK for macOS (Make sure to select "System Global Installation" during Select Components.)

Inspired by https://infosec.exchange/@jjtech/112612685494089718

Method 1: LLDB

Step 0

Reboot into Recovery OS + Disable SIP

csrutil disable

FAQ on the xz-utils backdoor (CVE-2024-3094)

This is a living document. Everything in this document is made in good faith of being accurate, but like I just said; we don't yet know everything about what's going on.

Update: I've disabled comments as of 2025-01-26 to avoid everyone having notifications for something a year on if someone wants to suggest a correction. Folks are free to email to suggest corrections still, of course.

Background

Displaying HDMI sources on Apple Vision Pro

While it's possible to stream most content to Apple Vision Pro directly over the internet, having the ability to use Apple Vision Pro as an HDMI display can still be useful.

Since Apple Vision Pro does not support connecting to an HDMI input directly or using an HDMI capture card, we have to be a little creative to make this work. NDI provides the ability to stream HDMI content over a local network with really low latency, and it works great with Apple Vision Pro.

This page shows the setup I’m using.

	import SwiftUI
	#if canImport(WidgetKit)
	import WidgetKit
	#endif

	/// Displays the app's icon in the current appearance (light, dark, clear, etc).
	///
	/// - note: The appearance of the icon will only reflect what's currently selected in System Settings,
	/// it will not update in SwiftUI previews or if the app is overriding its own appearance.
	public struct AppIconView: View {

	"""
	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

	// kdebug_interpose.c

	#include <stdbool.h>
	#include <stdint.h>
	#include <dlfcn.h>

	// Forward declare the original
	extern bool kdebug_is_enabled(uint32_t debugid);

	// Our replacement

	// Sometimes you have a large file on a small disk and would like to "transform"
	// it in some way: for example, by decompressing it. However, you might not have
	// enough space on disk to keep both the the compressed file and the
	// decompressed results. If the process can be done in a streaming fashion, it
	// would be nice if the file could be "drained"; that is, the file would be
	// sequentially deleted as it is consumed. At the start you'd have 100% of the
	// original file, somewhere in the middle you'd have about half of the original
	// file and half of your output, and by the end the original file will be gone
	// and you'll be left with just the results. If you do it this way, you might
	// be able to do the entire operation without extra space!

	// UserDefaults KVO observation with AsyncSequence/AsyncStream
	// Ole Begemann, 2023-04
	// Updated for Swift 6, 2024-11
	// https://gist.github.com/ole/fc5c1f4c763d28d9ba70940512e81916

	import Foundation

	// This is ugly, but UserDefaults is documented to be thread-safe, so this
	// should be OK.
	extension UserDefaults: @retroactive @unchecked Sendable {}