← All games

The LLM Pavilion

A model that writes essays, at a board it cannot hold

The frontier moved, the border stayed jagged

Most of this site documents an old campaign: minimax, alpha-beta, Monte Carlo tree search — machinery built to win games, doing exactly that, one game at a time, for seventy years. By 2026 the attention has moved elsewhere. The machines everyone talks about now write essays, hold conversations, pass professional examinations — and lose board games to a search algorithm old enough to have grandchildren.

There is no paradox in this, only a difference of construction. A language model is trained to continue text: given everything so far, produce a plausible next piece. Nothing in that objective builds a board, tracks which square is occupied by what, or checks a line of play ten moves deep. A search algorithm does nothing else. Where the two kinds of machine meet, the border between them is as jagged as the one this site has always documented between machines and people — hard and easy have simply changed places again.

This page is that border, shown rather than argued. Below: two real games between a language model and this site's own Connect Four engine, recorded move by move; a quiz built from the model's other famous weakness; and an honest, dated map of where the seams currently run. Everything on this page was computed before it reached you — it works offline, and nothing here phones anywhere.

The model at the board

Two genuine games of Connect Four, played in August 2026. The blue discs were chosen by Claude, a language model, reasoning over a text drawing of the position — no search, no lookahead beyond what fits in a paragraph of thought. The red discs were computed by this site's expert engine from the Connect Four page: plain alpha-beta search, the forty-year-old kind. Step through the moves; the commentary under the board is the model's own, and it is not kind to itself.

Claude (language model)site engine (search)
Move 0 of 28
Start of game 1. Claude moves first. Press "forward".

Score of the match: engine 2, model 0. That is the finding, reported as it happened. The model saw many single threats in time, reasoned plausibly about squares and windows — and lost both games to double threats and zugzwang, the things you find by exhaustive search and miss by plausible reasoning. On this site's difficulty ladder the language model sits somewhere below the easy level of a 1970s program.

Both games: Claude (Anthropic) against the expert level of this site's connect4.html engine, August 2026. The engine code was extracted unchanged; every recorded move was re-verified for legality by replaying the full game through the engine. The games are reproduced in full — nothing edited, nothing retried.

Spot the fabrication

The model's other famous failure is confabulation: a wrong detail delivered in the same confident prose as a right one. Below are five rounds of three short passages about the history told on this site. In each round two passages are accurate and one contains a plausible invented detail. Tap the one you think is invented. Your score stays on this device.

Round 1 of 5
A.Nim was the first board game to be solved: in 1901 Charles Bouton published the complete winning strategy — on paper, half a century before computers existed.
B.Connect Four on the 7×6 board was solved in 1988 by Victor Allis and James Allen, independently of each other: with perfect play the first player wins.
C.English draughts were solved by Jonathan Schaeffer's Chinook team in 2007, after eighteen years of computation: with perfect play on both sides the first player wins.
All fifteen passages were written by Claude (Anthropic), August 2026. The true details follow this site's reference pages — Solved Games, Landmark Matches, the timeline — and the sources cited there. Each fabrication is marked in this page's data and named in the reveal; nothing is generated at run time.

Where the models stumble (2026)

A dated list, because it will age. This is where the seams ran in August 2026 — each entry on the first list is a thing a cheap classical program does flawlessly, or a careful person does reliably, while a language model does not.

Weaker than a good tool, and often weaker than you

And the mirror list — stronger than most people

Whether you can tell the model's prose from a person's is a separate experiment, and this site already runs it: Spot the Machine. Expect to lose — and note which direction that boundary moved, and how fast.

How this page was made

The games and the texts were generated by the language model Claude (Anthropic), August 2026. In the games, Claude chose every blue move by reasoning over a text drawing of the board, with no game-tree search assisting the choice; the opponent was the unmodified expert level of this site's own Connect Four engine. The finished games were replayed through the engine to verify legality, and are shown unedited, including the losses. The quiz passages were written by the same model; their true details are checked against this site's reference pages, and every fabrication is deliberate, marked in the page's data and disclosed in the reveal. No content on this page is fetched, generated live, or attributed to any system that did not produce it.

Games and texts generated by the Claude language model (Anthropic), August 2026. Engine: connect4.html, expert level, extracted verbatim. Facts checked against solved.html, matches.html and timeline.html and the sources cited there.
Where to go from here
To cite this page: “The LLM Pavilion”, You vs Computer, https://youvscomputer.com/games/llm.html