Journal
Posts and long-form series on AI, startups, venture capital, and more.
All Posts
Ruin Is a Result, Not an Exception
If your simulation can end early, the ending is part of what it returns. Where to raise, where to catch, and why the difference decides what you can analyse.
What a Strategy Needs to Be Told About a Settled Bet
A staking strategy that cannot see outcomes is static whatever the class is called. The callback that fixes that has three decisions worth making deliberately.
What Fractional Kelly Actually Buys You
Betting half of Kelly costs about a quarter of your growth. Betting double Kelly costs all of it. The curve is symmetric and the risk is not.
Nine Bankroll Strategies Under Four Kinds of Stress
keeks v0.6.0 ships a deterministic risk benchmark. Nine staking strategies, four ways to make life hard, and the one that matters most.
When the Fee Eats the Edge
A flat fee does not scale with your bet, and Kelly sizes small when the edge is thin. The interaction costs more than either of them looks like it should.
Alloy Composition Optimization: Multi-Objective Materials Design
Using genetic algorithms to optimize alloy compositions for multiple properties simultaneously, balancing strength, hardness, density, and corrosion resistance.
Rocket Motor Optimization: Designing Thrust Profiles
Using genetic algorithms to optimize solid rocket motor grain geometry for target thrust profiles, where physics simulations replace analytical gradients.
Twelve Rating Systems, Walked Forward Through Eight Seasons
Do rating systems built for college football beat ones built for chess at predicting college football? Six thousand games, one week at a time.
HP Lattice Folding: Simplified Protein Structure Prediction
Using genetic algorithms to fold proteins on a lattice model, where the goal is to maximize hydrophobic contacts in a discrete search space.
Scoring a Three-Outcome Predictor
A rating system gives you one number, and a draw is a third outcome that number cannot express. What that does to your accuracy, measured on 40,000 chess games.