Word Freak AI

Teaching machines to play the world's greatest word game

Scroll

The Problem

Every major board game has had its public AI moment — the kind the world stops to watch. Chess had Deep Blue. Go had AlphaGo. Poker had Pluribus. S1C3R1A1B3B3L1E1 has incredible engines — but it's never had that moment.

Meanwhile, the frontier AI models that pass the bar exam and write working code? They can barely play Scrabble at a beginner level. They can't count the letters in their own input. They can't read a board. They confidently play words that don't E1X8I1S1T1.

Frontier LLMs at Scrabble
~800
Estimated rating. Below casual club players.
Best Engine on Earth
~2,400
BestBot. Beats top humans 60% of the time.
The Human GOAT
~2,200
Nigel Richards. Won the French Worlds without speaking French.
Word Freak's Current Bar
<55%
Still chasing a consistent win rate against Quackle.

The Generations

Word Freak doesn't exist in a vacuum. Thirty years of brilliant work got us here. Every generation solved problems the last one couldn't — and each one made the next possible.

Gen 1

Maven

1986 – 2000s
Brian Sheppard

The first serious Scrabble AI. Sheppard's PhD work proved that simulation-based move evaluation could beat the best humans. Maven defeated world champion Adam Logan in a 1998 exhibition match and showed the world that computers could master word games, not just chess.

Proved it was possible
Gen 2

Quackle

2005 – present
Jason Katz-Brown & John O'Laughlin

The open-source engine that became the community standard. For nearly twenty years, "check it against Quackle" was how competitive players studied. Its Monte Carlo simulations and hand-tuned heuristics set the bar for what a Scrabble engine could be. Still the reference point today.

Gave the tools to the community
Gen 4

Word Freak

2025 –
Rounds for Squares

We're trying something different: pure self-play reinforcement learning. No human games, no hand-crafted heuristics. A neural network that teaches itself by playing millions of games against itself — the same approach that produced AlphaGo. We're also building a native mobile experience for competitive word games, because players deserve a great app in their pocket. We'd love to build it together.

Standing on shoulders

Everything we're building exists because of the work above. The Woogles team built the platform the community needed. We think there's room to grow the game together — especially on mobile, where competitive players still don't have a real home.

What We're Building

Word Freak combines a high-performance Rust engine — finding every legal M3O1V4E1 on a board in under three milliseconds — with a neural network trained through millions of games of self-play. The same technique that produced AlphaGo, applied to the world's most popular word game.

We're also building a benchmark that exposes where frontier AI breaks, a coaching tool that helps competitive players improve, and a narrative that ends with the greatest player who ever lived facing a machine that taught itself to play.

System Type Rating Status
BestBot Engine · CNN + Monte Carlo · Woogles ~2,431 The standard
Nigel Richards Human ~2,200 The GOAT
Quackle Engine · Classic ~2,100 Legacy standard
GPT / Claude / Gemini LLM ~800 Can't count letters
Word Freak Engine · Neural + RL TBD 1 BestBot win; still below 55% vs Quackle

The Goal

“He memorized the entire French Scrabble dictionary — 386,000 words — in nine weeks. He doesn't speak French. Then he W4O1N1 the World Championship.” — On Nigel Richards, the greatest Scrabble player who ever lived

Word Freak's ultimate goal is a formal match against Nigel Richards — Scrabble's AlphaGo moment. A best-of-fifty series between a machine that taught itself to play and the most dominant competitor in any game, anywhere.

But first, we have to beat Quackle consistently. Word Freak has one BestBot win, but the repeatable benchmark still says we're below the level we need.

The Journey

Complete

The Engine

Rust GADDAG engine. Every legal move in under 3ms. 1,500+ games per second.

Complete

The Benchmark

ScrabbleBench 2.0. Four tiers testing vocabulary, spatial reasoning, strategy, and full gameplay across four frontier models.

Complete

The Neural Network

Self-play training in 25,000-game milestones. Each generation challenges the champion. Only the strong survive.

Complete

First BestBot Win

Word Freak won one game against BestBot in April 2026. Promising, but probably high variance.

In progress

Crack Quackle

We still have not reached a 55% win rate against Quackle. That is the next honest milestone.


By the Numbers

Word Freak Agent-0 is in its sixth generation, learning entirely through self-play. No human games. No hand-crafted heuristics. Just millions of positions, each one teaching the network a little more about what makes a great move.

Agent Version
v6
Sixth generation. Pure self-play evolution.
Games Played
11.4K
And counting. Every game teaches the next.
Positions Analyzed
961K
Each with every legal move scored and ranked.
Neural Features
51
Leave quality, premiums, geometry, defense, bingo potential.
Win Rate vs Greedy
56%
WF Agent-0 v4 vs. a pure score-maximizing baseline.
Move Generation
<3ms
Every legal move on any board. Rust GADDAG engine.

Game Highlights

When you play thousands of games, extraordinary things happen. These are real moments from WF Agent-0's self-play training — automatically extracted from the games as they're played.

Triple-Triple Bingo
AERIFIED
194 pts
All 7 tiles played across two triple-word-score squares. A 9x word score multiplier plus the 50-point B3I1N1G2O1 bonus. Swung the lead by 194 points in a single turn.
Photo Finish
357 – 357
Exact tie
In 27 turns of play, both agents finished with the exact same score. Three games out of 10,000 ended in a perfect tie.
Comeback Win
128-point deficit
overcome
Down by 128 points — nearly half the final total — WF Agent-0 clawed back to win 411–401.
6-Way Crossword
EXITED
6 words formed
A single move that created six perpendicular words simultaneously. Every tile placed formed a valid crossword with the existing board.
Blowout
613 – 268
345-pt margin
Total domination. A 613-point game is elite by any standard — then winning by 345 on top of it.
Unusual Word
PREALLOTTING
11 letters
Not a word most humans would find. WF Agent-0's GADDAG engine sees every legal possibility — including the obscure ones.

Highlights extracted from 10,000 self-play games. As training scales to millions, the moments only get better.


Latest Updates

All Updates

Follow the Journey

Word Freak is being built in public. Watch the engine L1E1A1R1N1, see the benchmarks update, and follow the road to the match.