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.

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.

Some notes, tools, and techniques for reverse engineering macOS binaries.

• Reverse Engineering Tools
  • Binary Ninja
• Ghidra

Current Status (as of December 24, 2025):
Discord officially removed the HypeSquad section from settings in September 2025, meaning you can no longer join via the in-app quiz. Existing badges remain visible on profiles, but new users can't join officially.

Good news: The internal API endpoint (/hypesquad/online) still works! This console script lets you add a HypeSquad house badge directly.

Important Warnings:

• This is unofficial and may violate Discord's Terms of Service (considered automation).
• Use at your own risk—account bans are rare for this but possible.

Cross-variant comparison of all registers read by tools/mmio-probe/probe.sh. 15 Ampere cards probed: 2× CMP 170HX 10GB (Cab + Own physical hardware, 2026-05-05/07) + 11 via Vast.ai rentals + 2× Drive A100 32GB (PG199, GA100-550F-A1, physical hardware, 2026-05-31).

• Key findings
• FMA / Tensor Core throttle
• Feature override chain

Cross-variant comparison of VBIOS structure and configuration across the GA100 product family. Sourced from static binary analysis of ROM dumps + empirical flash experiments + Lapsus Booter disassembly.

Last updated: 2026-05-31 (Drive A100 PG199 VBIOS dumped and analyzed)

Contributors: Petri Krohn (ECB cryptanalysis, RFRD manifest decode, known-plaintext identification), Cab (ECB padding block confirmation, ImHex structure labeling, license region offset discovery via gpuio BAR0 dump)

• Docker Desktop (or Docker Engine + Docker Compose v2)
• Ports 3000 and 8000 free on your machine

1. Create a folder and open a terminal there:

The following are examples of the four types rate limiters discussed in the accompanying blog post. In the examples below I've used pseudocode-like Ruby, so if you're unfamiliar with Ruby you should be able to easily translate this approach to other languages. Complete examples in Ruby are also provided later in this gist.

In most cases you'll want all these examples to be classes, but I've used simple functions here to keep the code samples brief.

This uses a basic token bucket algorithm and relies on the fact that Redis scripts execute atomically. No other operations can run between fetching the count and writing the new count.