Essays, research notes, and deep dives

•AI Agents & Infrastructure

Using AI Agents and Skills to Migrate Off Slack

Two paired Claude Code / Codex skills that move a company from Slack to self-hosted Mattermost — fail-closed readiness gate, named rollback owner, ~99% lower ongoing cost.

•AI & Prompting

The Overprompting Trap

Why giving AI models too many constraints degrades output quality. The chef analogy, the two-phase approach, and how to get out of the model's way.

•Statistics

My Favorite Statistical Measure: Hoeffding's D

Standard correlation misses non-linear relationships. This guide explains Hoeffding's D, a powerful non-parametric measure that detects complex dependencies where Pearson and Spearman fail.

•Investing

Factor Risk Models and the Hedge Fund Business

An insider's look at how 'smart' risk models like Barra often unknowingly distort incentives, encourage crowding, and create hidden systemic risks within multi-manager platforms.

•Algorithms & Networking

RaptorQ: The Black Magic of Liquid Data

A deep dive into RaptorQ (RFC 6330), the fountain code that turns any file into an infinite stream of interchangeable packets with 0.02% overhead.

•Algorithms

The Incredible Magic of CMA-ES

An interactive deep dive into the Covariance Matrix Adaptation Evolution Strategy. Explains how it adapts to high-dimensional, non-convex landscapes where gradient descent fails.

•Markets & AI

The Short Case for Nvidia Stock

A 12,000-word deep dive into how AI economics, models like DeepSeek, and GPU supply can collide with valuation narratives. Explores the potential reflexivity of the AI capex cycle.

•History & Science

The Most Impressive Prediction of All Time

Analyzes Pyotr Durnovo's 1914 memorandum to Tsar Nicholas II—an eerily accurate forecast of WWI, the exact alignment of powers, and the inevitable Russian Revolution that would follow.

•History of Math

The Lessons of Hermann Grassmann

The story of the self-taught genius who invented linear algebra (the wedge product) decades before it was understood. A lesson on how radical abstractions are often rejected by the establishment.

Building a Brain, Not a Calculator: The Bio-Inspired Nanochat Architecture

Transformers are static crystals; brains are living fluids. A deep dive into replacing weights with metabolic, fatigue-prone, and evolving biological analogs.

•Frontier Research

11 Ways to Break the Transformer: A Model-Guided Mathematical Odyssey

What happens when you let GPT-5 design its own replacement? An exploration of Lie groups, p-adic numbers, and tropical geometry in Deep Learning.

•Algorithms

Lamport's Bakery Algorithm

A visual, Pythonic implementation of Leslie Lamport's classic solution to the mutual exclusion problem. Illustrates how fair concurrency can be achieved without atomic hardware primitives.

•AI Research

LLM Introspective Compression

Proposes treating LLM context as a 'save state'. By allowing models to compress and manipulate their own internal thought states, we unlock capabilities like reasoning backtracking and reinforcement learning.

•AI Research

Multi-Round LLM Coding Tournament

Demonstrates that collective intelligence outperforms individual genius. A framework where multiple LLMs iteratively refine coding solutions via peer review to break out of local optima.

•Dev Workflow

Making Complex Code Changes with Claude Code

Advocates for 'separating cognition' when using agents. Instead of asking for code directly, use agents to generate granular plans first, spreading the reasoning load across more tokens.

•Security

Protecting Against AI Prompt Injection

Traces the history of 'jailbreaking' from simple commands to recursive meta-prompts. Argues that internal guardrails always fail and proposes external 'inoculation' strategies.

•Healthcare & AI

Dr. GPT: Empowering Your Healthcare Decisions

How to use AI as a tireless patient advocate. Describes a method for synthesizing scattered medical records into a unified dossier to catch missed diagnoses and dangerous drug interactions.

•Utility

TaxGPT: Using AI for Tax Prep

A practical guide to decomposing tax returns into context-sized chunks for AI analysis. How to find deductions, spot errors, and out-perform human accountants on complex personal returns.

•Software Architecture

Engineering the Mindmap Generator

Why pipelines aren't enough. Details the 'non-linear exploration model' architecture needed to extract deep, hierarchical structure from complex documents without hallucination.

•Data Science

PPP Loan Fraud: A Data Science Detective Story

Reconstructs massive fraud patterns in the PPP loan program. Shows how network analysis and simple heuristics could have caught billions in theft that standard banking checks missed.

•Engineering

Building the Python Backend for YTO

A technical retrospective on using FastAPI, SQLModel, and Pydantic to build a heavy-duty async backend. Covers Whisper integration and optimizing for concurrency.

•Web Development

Introducing Our New Next.js Blogging System with Tailwind CSS

A look at our sleek and efficient new blogging system built with Next.js and Tailwind CSS, designed for fast, responsive, and visually appealing content delivery.

•AI Safety

Some Thoughts on AI Alignment

Argues that internal alignment (Asimov's laws) is a losing battle. Proposes an external 'criminal justice system' for AI: helper models that monitor and police the primary model's output.

•Web Dev

Next.js GitHub Markdown Blog System

Explains the architecture behind the open-source blogging engine. How to treat GitHub as a headless CMS, using static generation for speed and Git for version control.