Building AlphaGo from scratch – Eric Jang

| Podcasts | May 15, 2026 | 93.7 Thousand views | 2:37:18

TL;DR

Eric Jang demonstrates how modern LLM coding tools and algorithmic improvements have democratized AI research, enabling a single researcher to rebuild AlphaGo for thousands of dollars rather than millions, while explaining how Monte Carlo Tree Search combined with neural networks solved a game previously considered computationally intractable.

🎯 The AlphaGo Revival 3 insights

LLM Coding Democratization

Thanks to modern coding assistants, recreating AlphaGo now costs thousands in compute rather than millions in research funding and team resources.

Personal Research Motivation

Eric pursued this project to understand how shallow ten-layer neural networks can amortize the simulation of extremely deep game trees.

KataGo Efficiency Breakthrough

David Wu's open-source KataGo (2020) achieved a 40x reduction in compute needed to train a strong Go bot tabula rasa compared to AlphaGo Zero.

Go Fundamentals 3 insights

Simple Rules, Complex Play

Players alternate placing black and white stones to surround territory and capture opponent stones by occupying all four adjacent intersections.

Tromp-Taylor Algorithmic Scoring

Unlike human scoring requiring consensus on dead stones, Tromp-Taylor rules provide completely unambiguous endgame scoring ideal for computer implementation.

Combinatorial Explosion

A 19x19 board allows roughly 361^300 possible game sequences, exceeding the number of atoms in the universe and making exhaustive search impossible.

🔍 Monte Carlo Tree Search 3 insights

PUCT Action Selection

The algorithm selects moves by maximizing the sum of mean action value (Q) and an exploration bonus weighted by prior probability and visit counts.

Efficient State Representation

Nodes represent game states storing visit counts, action values, and prior probabilities, enabling efficient tree traversal without storing the entire game tree.

Stochastic Approximation

Though Go is deterministic, Monte Carlo methods introduce probability distributions to sample promising game trees rather than exhaustively searching all possibilities.

Bottom Line

Modern LLM tooling has collapsed the implementation cost of complex AI systems like AlphaGo from millions to thousands of dollars, democratizing access to frontier research techniques that were previously exclusive to well-funded labs.

More from Dwarkesh Patel

View all
The math behind how LLMs are trained and served – Reiner Pope
2:13:41
Dwarkesh Patel Dwarkesh Patel

The math behind how LLMs are trained and served – Reiner Pope

Reiner Pope explains the mathematical mechanics behind LLM inference costs, demonstrating how 'Fast Mode' APIs charge premiums for smaller batch sizes that reduce latency, and why physical memory bandwidth constraints create hard limits on how fast or cheap inference can get regardless of budget.

21 days ago · 9 points
Michael Nielsen – How science actually progresses
2:03:04
Dwarkesh Patel Dwarkesh Patel

Michael Nielsen – How science actually progresses

Michael Nielsen dismantles the pop-science narrative of linear scientific progress through crisp experiments, revealing instead a messy, decentralized process where mathematical formalism often precedes conceptual understanding, expertise can blind researchers to truth, and communities adopt paradigm shifts long before experimental closure.

about 1 month ago · 10 points