Monte Carlo Tree Search for Code Generation using LLMs [ github ]
We combine a pre-trained transformer model aka deepseek-coder with an adapted version of the Monte-Carlo Tree Search (MCTS) algorithm to guide the generation of code generation to solve HumanEval and VerilogEval.
GPTX: A simple implementation of GPT-2 in Rust and CPP on CPU [ github ]
We benchmark the multiprocessing capability of Rust and CPP by comparing simple implementations of GPT-2 using Rust+Rayon, and CPP+OpenMP libraries.
Three problems to motivate Bayesian Machine Learning
Andrew Wilson's course on Bayesian ML really sold me on the idea of probabilistic thinking. The idea of bayesian model average, infinite parameters and epistemic uncertainity. I write about the solving 3 problems using a bayesian framework.
Benchmaring Rust vs. CPP for graphics [ github ]
I write about implementing 3 graphics programs in Rust benchmarked with CPP, and observe that Rust is faster. Comparision only on CPU.
An intuitive guide to solving a Rubik's cube using only 1 algorithm
I explore the Old Pochman method to solve a Rubik's cube, often used to solve blindfold. I go on to show that Old Pochman can be done in only 1 algorithm.
Thoughts on building a theoretical Clash Royale AI
Clash Royale, a mobile phone card game, is 21st century chess. I write about how to formally model it as a computational problem, and solve it using Reinforcement Learning. I talk about what a Clash Royale AI should be able to achieve in the process of solving the game.