Suraj ghishadow

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.

UNIX famously uses fork+exec to create processes, a simple API that is nevertheless quite tricky to use correctly and that comes with a bunch of problems. The alternative, spawn, as used by VMS, Windows NT and recently POSIX, fixes many of these issues but it overly complex and makes it hard to add new features.

prepare() is a proposed API to simplify process creation. When calling prepare(), the current thread enters “preparation state.” That means, a nascent process is created and the current thread is moved to the context of this process, but without changing memory maps (this is similar to how vfork() works). Inside the nascent process, you can configure the environment as desired and then call prep_execve() to execute a new program. On success, prep_execve() leaves preparation state, moving the current thread back to the parent's process context and returns (!) the pid of the now grownup child. You can also use prep_exit() to abort the child without executing a new process, it similarly returns the pid

These are simplified examples I came up with to go along with William Woodruff's excellent blog post on x86-64 addressing modes.

All the examples are compiled with GCC -O3 or -Os, which is indicative of what the compiler thinks the best option is in practice.

Base + Index

Use case: Indexing into a byte array (whose address is not fixed)

char f(char buf[], int index)

This document explains how I would implement an SSA-based compiler if I was writing one today.

This document is intentionally opinionated. It just tells you how I would do it. This document is intended for anyone who has read about SSA and understands the concept, but is confused about how exactly to put it into practice. If you're that person, then I'm here to show you a way to do it that works well for me. If you're looking for a review of other ways to do it, I recommend this post.

My approach works well when implementing the compiler in any language that easily permits cyclic mutable data structures. I know from experience that it'll work great in C++, C#, or Java. The memory management of this approach is simple (and I'll explain it), so you won't have to stress about use after frees.

I like my approach because it leads to an ergonomic API by minimizing the amount of special cases you have to worry about. Most of the compiler is analyses and transformations ov

Why did you do this?

Relax, I only have one Sunday to work on idea, literally my weekend project. So I tried Deepseek to see if it can help. Surprisingly, it works and it saves me another weekend...

What is your setup?

Just chat.deepseek.com (cost = free) with prompts adapted from this gist.

Does it work in one-shot or I have to prompt it multiple times?

RISC-V Vector Extension for Integer Workloads: An Informal Gap Analysis

Note: To verify my RVI membership and idenity on this otherwise semi anonymous account: I'm Olaf Bernstein, you should be able to view my sig-vector profile, if you are a member of the vector SIG.

The goal of this document is to explore gaps in the current RISC-V Vector extensions (standard V, Zvbb, Zvbc, Zvkg, Zvkn, Zvks), and suggest instructions to fill these gaps. My focus lies on application class processors, with the expectation that suggested instructions would be suitable to become mandatory or optional instructions in future profiles.

I'll assume you are already familiar with RVV, if not, here is a great introduction and here the latest RISC-V ISA manual.

10x Optimization! An installer creation adventure

Background: creating an installer tool

Lately I've been working on rInstallFriendly v2.0, my simple and easy-to-use tool to create fancy software installers.

rInstallFriendly v2.0, some of its multiple visual styles available. Style is fully customizable!

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

MiniPerf

Example of how to capture CPU counters with ETW on Windows, perf on Linux or kperf on Apple.

Using ETW needs somewhat recently updated Windows 10 or 11. Not sure about exact version.

Currently tested on:

etw on Qualcomm Snapdragon X Elite, Windows 11, arm64
etw on AMD Zen 3, Windows 11 (with virtualization enabled in BIOS)

	"""
	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