Interactive website at https://sane24.github.io/NYT-Wordle-Solver-AI/ where you can play against the AI, watch evolution and the genetic algorithm live, get hints using information theory.
One puzzle, solved with two kinds of intelligence. This project solves Wordle with:
- An information-theory solver - picks the guess with the highest entropy: the one whose feedback pattern is expected to shrink the space of possible answers the most. It reasons in information bits.
- A genetic-algorithm solver - a population of 100+ agents, each with a genome of strategy weights (positional letter frequency, 50/50 split-seeking, when to stop probing and commit…). Every generation the agents play real Wordle games; the weakest half dies, elites survive, parents crossbreed, children mutate. No entropy math is hard-coded - good strategy evolves.
The Evolution Lab runs a real genetic algorithm in your browser. A population of ~100 agents plays actual Wordle games every generation; the weakest half dies, elites survive, parents crossbreed, children mutate. Below is generation 12 solving a shared target — every one of the 100 agents lands the word within four guesses. Ten generations earlier, most of them couldn't solve it at all.
Each dot in the population view is one agent, plotted by average guesses vs. solve rate. You can watch the cloud drift toward the top-left corner as better strategies take over:
Measured on the full 2,315-word answer list:
| Solver | Avg guesses | Solve rate |
|---|---|---|
| Entropy AI | 3.46 | 100% |
| Evolved GA champion | 3.64 | 99.3% |
| Random consistent guesser | ~5 | ~85% |
On human-picked words it's not even close: over 300 common words outside the official answer list (the kind people choose in custom games), the champion solves 94.7% — a champion evolved without the common-word gene managed 77.0%.
The gap between the two is the interesting number: evolution gets within ~0.2 guesses of the information-theoretic approach purely by selection pressure, with no entropy math anywhere in its genome.
For a guess G over a candidate set C, the 243 possible green/yellow/gray feedback patterns partition C into buckets. The expected information of the guess is
H(G) = log2|C| − (1/|C|) · Σ n_b · log2(n_b)
The solver plays the guess with the highest H, tie-breaking toward words that could themselves be the answer (they can win outright). This is why it opens with strange words like SOARE — it's not trying to guess the answer early, it's trying to learn the most. First-move entropies over the whole dictionary are precomputed so the site loads fast.
An agent's genome is just 8 numbers: weights for positional letter frequency, overall letter frequency, split-seeking (preferring letters that appear in ~50% of remaining candidates), unique letters, answer bias, and common-word bias — plus a commit threshold and commit turn that decide when to stop probing for information and start actually trying to win.
The common-word bias gene is backed by real English frequency data (Norvig's web-corpus unigram counts, baked into core/freqs.js): when several words fit the clues, an agent can prefer the one people actually use. And there's genuine selection pressure behind it — a slice of every generation's fitness games uses custom targets drawn from the full dictionary weighted by how common they are, the way humans pick words, so knowing MOUSE is likelier than MOUST is a survival trait, not a hand-coded rule.
Fitness is the penalty-adjusted average guess count over a fresh sample of real games each generation. Selection is tournament-based with elitism, uniform/blend crossover, Gaussian mutation, and a few random immigrants per generation to keep the gene pool from stagnating.
My favorite result: high split-seeking weights keep winning, generation after generation. That's the 50/50 partition idea at the heart of information theory — rediscovered by evolution, without anyone writing down a logarithm.
A solver assistant for your own games. Mirror any Wordle (including the real NYT one) by typing your guess and clicking the tiles to match the colors you got. The engine tells you how many answers remain, how many bits of uncertainty are left, and the best next guesses, ranked by expected information:
A head-to-head mode that races every solver on the same word and shows their full reasoning traces side by side. Both solvers nail MOUSE in 3 — the entropy AI through alien probes like SOARE and PILUM, the GA champion through everyday words like ALERT and NOISE — while the random baseline runs out of turns. It tells you more about the three strategies than any chart:
A Discord bot (bot/) that wraps the same solver core in slash commands: /wordle play to race the AI, /wordle hint for help with any position, /wordle solve, /wordle compare, and a /wordle daily puzzle.
For the Discord bot, copy bot/.env.example to bot/.env, fill in your bot token, then:
npm run bot:deploy # register the slash commands once
npm run botcore/ solver engine (JS, no DOM/Node deps)
feedback.js pattern encoding: 5 tiles → one byte, base 3
entropy.js information-theory solver
genetic.js genomes, fitness, selection, crossover, mutation
tables.js typed-array lookup tables for the hot path
daily.js deterministic daily word
js/ website: tab router, boards, workers
scripts/ offline evolve / benchmark / precompute
bot/ Discord bot (discord.js)
- Word lists are the public Wordle answer list (2,315 words) and legal-guess list (12,972). The daily word here is intentionally not the official NYT word, so nothing gets spoiled.
- Custom words outside the answer list work everywhere — the solvers detect it and fall back to the full legal-guess dictionary.