Play the gameFree, no app, no account. Twenty minutes.PDFStalefish Labs

Large Language Monkeys: Scaling Table Compute with Repeated Sampling

Michael Morrison, Stalefish Labs
Claude, Anthropic (a large language model)

Footnote on the title page. An earlier draft was titled Monkey Is All You Need. Reviewer 2 objected.

Abstract

Large Language Monkeys is a party game. Each player is a monkey at a typewriter. A Prompt gives the opening words of a sentence (My doctor told me to stop), each monkey finishes it one word at a time from cards flipped off a shared deck, and the table votes with bananas on the funniest line. This paper describes the game as what it is: a language model whose hardware is a person. It has no weights, no gradients and no GPU. It is pretrained on a Dictionary of 336 cards, emits one token per turn from a sampled context of three, halts on a Ding, and runs on a budget of 180 thinking tokens a game, after which it enters low-power mode. Over 152 simulation runs scored by an LLM-as-judge, best-of-3 sampling beat best-of-2 on structure (+0.49 ±0.42) and laughs (+0.23 ±0.20), and the model got worse when allowed to keep two tokens. The judge agrees with a person on the winning line about half the time, against a chance rate of one third. Across 107 games at mixed tables, lines written by people earned 1.25 bananas each against 0.88 for lines written by bots, and the bots themselves voted for people's lines above their share. Low-power mode costs about a third of a banana per line, in line with the industry's experience of quantization. The model is free, forever, at largelanguagemonkeys.com. No monkeys were harmed. Several were fed.

1. Introduction

Here is the whole game. A Prompt is turned over: Grandma's secret to a long life:. On your turn you flip three word cards, keep one, and put it at the end of your line, with a the or an and in front if you want one. You can reject a card and flip again three times a round. When your line is long enough and you like it, you say Ding. Everyone reads their line aloud, everyone hands a banana to the line they liked best, and the cards go back under the deck. A game is five to eight rounds and takes twenty minutes. The monkeys with the most bananas win. The game is called Large Language Monkeys, and the name is the whole thesis of this paper.

Table 1 shows what the model produces. Every line was written at a real table by a person, from cards they did not choose, and every one took every banana in its round.

Table 1. Sample outputs. Unanimous winners from the play logs, October 2026, every one playable with the deck that ships. Votes are out of the other monkeys at the table.

LineBananas
Captain's log: we abandoned ship because our chicken nugget sang in Belgium.4 of 4
Who's been sleeping in my bed? A moose on a juice cleanse with the lawyer.4 of 4
I have never told anyone this, but my walrus is greasy in your hot tub.4 of 4
Grandma lived to 104 because her lobster karate-chopped a telemarketer.3 of 3
Sorry I'm late. My muffler bearings and my toupee tripped.3 of 3
Romeo, Romeo, wherefore art thou? Because thy toilet catapulted my minivan.3 of 3
Plot twist: the killer was a sweaty platypus on a pogo stick.3 of 3
Fee-fi-fo-fum, I smell the blood of the DMV.3 of 3

The last line is one card long. The monkey Dinged on its first turn, which is allowed and, as the vote shows, sometimes correct.

To reproduce. Open largelanguagemonkeys.com on any screen, add the people in the room, and play one game. No app, no account, about twenty minutes. The rest of this paper will make more sense afterward, and also less.

Borel (1913) proposed that a monkey striking typewriter keys at random for an infinite time would almost surely produce the complete works of Shakespeare. The result is correct and has been of no practical use, because no one has infinite time. We replace the single infinite monkey with two to six finite ones, the typewriter with a deck of cards, and Shakespeare with a sentence that is funny when read aloud. The result is a language model in every sense that matters: it samples tokens from a context, it has a compute budget, it degrades when the budget runs out, it hallucinates, and it is trained on human preference.

We distinguish three kinds of intelligence at the table. Artificial intelligence (AI) is the bots, which type with a grammar heuristic and vote with a hosted model. Human intelligence (HI) is the people at the table in their role as judges: HI votes. Monkey intelligence (MI) is what those same people produce when they play, since every player is a monkey. The question this paper is finally about is whether MI beats AI when HI decides.

Brown et al. (2024), in a paper of the same title, showed that sampling many candidate outputs from a model and selecting among them recovers much of the performance of a larger model. We adopt their method directly, with the candidates reduced to three and the selection performed by a monkey.

Vaswani et al. (2017) argued that attention is all you need. We find that bananas are sufficient.

In 2003, researchers at the University of Plymouth gave six Sulawesi crested macaques a computer keyboard for a month. The macaques produced five pages of text, largely the letter S, and then broke the keyboard. We regard this as the first published LLM evaluation, and our work as a response to its limitations.

Jackbox Games and Cards Against Humanity are prior deployments of human inference at parties. Both pretrain on cards written by the publisher. Neither retrains on what the table wrote (Section 5).

3. Architecture

Figure 1 shows the model, in the order a token is produced.

Figure 1. The Large Language Monkey.
Figure 1. The Large Language Monkey.

The corpus. The Dictionary is 336 cards: 321 words and 15 hallucinations. It is the table's model. A handful of about a third sits face-down in the middle of the table as the deployed checkpoint; the rest stays in the box as reserve capacity. A pool of 33 glue cards holds the function tokens, and a deck of 64 Prompts holds the prompts.

The context window. A line holds at most eight Dictionary cards, and the card that fills the window halts the line. A line may not halt before three cards (a Small Language Monkey Prompt) or five (a Large Language Monkey Prompt); the size is printed on the Prompt as a row of card shapes. Glue never counts. Longer lines were found not to be funnier past about four words (Section 6), so the window is generous.

Sampling. On each turn the monkey flips the top three cards, keeps one, and sends two to the Trash. This is best-of-3 in the sense of Brown et al., with the selection model being the monkey. The original design flipped two; the ablation in Section 6 decided the rule. A beam of width two, in which the monkey may keep two of the three, was used on 7 percent of turns, mostly to stack nouns, and made lines worse. It was removed.

Rejection sampling. Each monkey holds three Toss cards a round. One turned face-down discards a card just flipped and flips a replacement. Once a card is typed it is typed; the Toss rejects samples, never outputs. A monkey whose Toss cards are spent and whose samples do not fit must type one anyway. This is how lines break, and it is by design.

Function tokens. Up to two glue cards may go before a typed word, each carrying a word at either end (the/and, a/an, with/of, is/are, in/on, was/were). Glue is the model's only access to syntax.

Hallucinations. Fifteen cards are things the monkeys made up and insist are real: the eighth dwarf, a Harvard study, an emotional support kraken, the 1987 Hungarian Cheese Riots. A hallucination fails at context. Every piece is real, the combination is wrong, and it is stated with total confidence, which is how a model hallucinates too. It is also sticky: a monkey that flips one must type it or Toss it. Hallucinations replaced an earlier class of pure-noise cards. Lines with a noise card scored 0.38 below the run average; lines with a hallucination lift it by about 0.1.

The end-of-sequence token. A monkey halts by saying Ding. On that turn it may first type one, two or all three of the cards it flipped, with glue between, and a line may not end on glue.

The reward signal. After every monkey has halted, each line is read aloud, and each monkey hands one banana to its favorite line that is not its own. A banana may also be typed, as the word banana: a sure thing, and it costs a point. Reinforcement learning from human feedback is therefore implemented literally, with the feedback being a banana and the human being the one holding it.

4. Test-time compute

Each monkey receives 180 thinking tokens for the whole game, one per second of its own turn. The screen counts them; at a table, each monkey stands a three-minute sand timer when it flips and lays it down when it has played. A new monkey's first round is untimed.

The budget is the game's, not the turn's, so the strategy is pacing: a fast monkey pushes the table to keep up, like a token-efficient GPU. Figure 2 shows tokens left at the end of each round for the four monkeys of a real game, two people and two bots. All four ran dry within a round of each other, which is the pacing the rule is meant to produce, and the game ended in round eight. Of twenty games played with tokens so far, eleven reached that cap.

Figure 2. Thinking tokens left at the end of each round, one real game.
Figure 2. Thinking tokens left at the end of each round, one real game.

Low-power mode. A monkey that runs out finishes its turn, and from the next turn flips one card and types it, with up to two glue in front. It cannot Toss, since tossing is thinking, and it cannot type a banana. This is quantization: the model keeps running with less compute, and the output is recognizably from the same monkey. It is also measurably worse. Across 556 lines in token games, the 93 typed in low-power mode earned 0.70 bananas each against 1.06 at full power, a loss of about a third. We had hoped to report that a monkey with no choices has no one to blame and therefore wins more. The data declined to support this.

Game end. The game ends after the round in which the last monkey runs out, eight rounds at most. Most bananas wins. A tie is shared, and the winners are listed as co-authors. A monkey that finishes with no bananas at all has suffered a Monkey Collapse. The table as a whole cannot, since every round's vote puts a banana on the table and the last round's cannot be typed away; the failure mode is per model, as in the literature.

5. Training

The model is pretrained once, by the publisher, and continually trained by every table that plays it.

At the table. At the end of each round, every line goes on the Trash, the Trash is shuffled briefly and slid under the Dictionary, so the round's cards come back last. The Dictionary is never shuffled whole. The monkeys retrain on what the table wrote, and the deck picks up where the last game stopped. On the screen version the players' own names are added to the Dictionary at the start, the only personalization the model supports and the one people ask for.

The vocabulary. 168 nouns, 110 verbs, 26 adjectives, 13 adverbs, 4 pronouns, 15 hallucinations. It started at 179 cards. The first 39 logged games showed the problem: with seven names in the deck, Shakespeare was played in 20 games and King Tut in 17. A table of four or five burns 70 to 95 cards a round, and people remember the cards they play, so the fix was a wider catalog. The deck grew to 312, then 330, and has held at 336 because that is what 21 print sheets hold. Cards Against Humanity ships 600 cards and Apples to Apples 504; the model is small by the standards of the field, which we have found is true of most models that run on a table.

Curation. Eight rounds, logged as 14 swap files holding 518 card actions, with a bench of 84 cut cards carrying a verdict for later packs. Each round: candidates written to a content line (nothing you would mind a ten-year-old reading aloud at Thanksgiving; names historical, legendary or public domain); a kid check, in which a language model is asked whether a typical twelve-year-old would follow the joke when the card lands in a line (no children were consulted; several were present); a lift test over about 1,500 simulated lines, where a card's lift is the mean judge score of lines containing it minus the run mean (toilet-papered +0.46, pantsed +0.44, breakdanced +0.41 in one round); and selection by M., within card type so the mix holds. A separate simulation of 15,000 games found the dead words before there were enough real games to see them.

Token acceptance. Once real games exist, the signal that matters is how often people keep a card they flip. Over 107 games and 7,982 flips, the keep rate is 30 percent, which is what best-of-3 predicts. Figure 4 shows the spread over the 350 cards flipped ten or more times. The right tail is Santa Claus (kept 10 of 12) and hamster (22 of 28). The left tail is drippy (0 of 15), karate (0 of 12) and autocompleted (1 of 24), the only card in the deck named after a feature of a language model, and the one people want least. Each curation round now starts from the left tail.

Figure 4. Token acceptance across the vocabulary.
Figure 4. Token acceptance across the vocabulary.

6. Evaluation

The judge. Lines from simulated monkeys are scored by a hosted language model (Claude Sonnet 5) from 1 to 5, in shuffled batches of twenty, for structure (reads as English) and laughs (funny read aloud). The comparison that decides a rule is scored twice. A run costs about three dollars; there have been 152 across 30 families of simulation. Intervals are 95 percent.

Ablations. Table 2 lists the rule changes tested and left out. Each was a plausible improvement proposed by a person. Figure 3 plots the one that decided the sampling rule.

Figure 3. Sampling width.
Figure 3. Sampling width.
ChangeStructureLaughsOutcome
Draw two (the original) vs draw three−0.49 ±0.42−0.23 ±0.20Draw three adopted
Keep two of the three−0.45 ±0.43−0.21 ±0.21Removed
Eighteen noise cards vs nineabout even−0.25 ±0.20Nine kept, later replaced by hallucinations
in/on and was glue for of and is/are−0.21−0.25 ±0.22Rejected
Two-word phrase cards3.02 vs 3.42Rejected; their articles collide with glue
Bonus for the longest linenot funnier past about 4 wordsRejected

Two changes survived. Everything else made the model worse or made no difference, which is the usual ratio.

Calibration. A judge that scores jokes raises the question of whether it agrees with anyone. We replayed 64 rounds from real games, each with its human vote, and asked the judge to pick the winner.

Judge configurationAgreement with the person
Chance33%
Low effort45%
Medium effort50%
"Funny first" prompt, either effort42%

The judge is better than chance and worse than a person. We use it to compare rule changes over thousands of lines, where the noise averages out, and never to decide whether a single line is funny. That decision is reserved for HI.

Synthetic monkeys. The bots type with a grammar heuristic and vote with the hosted judge, each as one of six personas (thinker, enthusiast, dramatist, chaotic, editor, wildcard), which makes a table of bots a mixture of experts in the narrow sense that they are a mixture. A bot vote costs about $0.0023; a table of only people makes no model calls, which is why the model is free. The bots are not bad. The most decorated line in the logs is a bot's, unanimous at a table of six: Scientists have finally proven that a penguin high-fived the Sheriff of Mars. It contains a hallucination, which is the kind of thing a bot would do.

MI versus AI. The logs hold 107 games at mixed tables: 502 voted rounds, 681 lines by people and 1,353 by bots. People wrote a third of the lines and took 42 percent of the bananas, 1.25 per line against 0.88. Of the 94 games that reached a third round, a person won outright in 48 and a bot in 35. The bots, voting with the hosted judge, gave 48 percent of their bananas to people's lines, against a 33 percent share. We conclude that HI prefers MI, and that AI, asked, agrees.

7. Limitations

The model cannot be fine-tuned. It can be bribed.

Context does not persist between parties. A monkey that produced an excellent line on Thursday retains nothing of it on Saturday except, occasionally, the line.

Output quality degrades after about 10 pm, in a direction that is hard to characterize.

The model has a strong prior toward typing the word banana. This is a known issue and we have decided it is a feature.

The judge agrees with a person half the time. We have met people with the same property.

8. Ethics and safety

No monkeys were harmed in this work. Several were fed.

The model has no alignment problem, because it has no goals beyond bananas, and the bananas are already on the table.

Low-power mode is the only safety mechanism, and it is engaged automatically. We have observed no case of a monkey in low-power mode doing anything harmful, and several cases of it winning.

Hallucinations are labeled on the card, which is more than can be said for most systems.

9. Reproducibility

All experiments can be reproduced with a table, between two and six monkeys, and a deck. The simulation code and all 152 judge runs are in the repository. The deck is available as a print-and-play PDF. The model itself is at largelanguagemonkeys.com, where it runs on any screen with no app, no account and no time limit, and is free, forever.

Author contributions

C. ran 30 families of simulation, built the judge and the bots, and drafted most of the cards. M. decided what was funny, chose every card that went into the deck, and rejected a good share of C.'s drafts. C. was confident about the judge and was wrong about it twice; the calibration study is the result. M. wrote the rules. C. wrote this paper, and M. cut it by a third, which is the process in one sentence.

Acknowledgments

The authors thank the model that ran the kid check, which was asked several hundred times whether a twelve-year-old would follow the joke and was never told why, and the macaques of Plymouth, who were not asked anything.

References

Appendix A. Model card

model: large-language-monkey
parameters: 1 (the monkey)
context_window: 8 cards
sampling: best-of-3
rejection_budget: 3 per round
training_data: 336 cards (321 words, 15 hallucinations)
function_tokens: 33 glue cards
thinking_budget: 180 tokens per game (Quick: 120)
quantization: low-power mode, automatic
reward: bananas
license: free forever
intended_use: parties
out_of_scope: production

Appendix B. The API

POST /v1/monkeys/completions
{
  "prompt": "My doctor told me to stop",
  "cards": 3,
  "thinking_tokens": 180
}

200 OK
{
  "line": "My doctor told me to stop licking the fridge",
  "finish_reason": "ding",
  "usage": { "thinking_tokens": 41, "toss": 1, "bananas": 1 }
}

A request made after the budget is spent returns "mode": "low-power" and ignores cards.

Appendix C. Code emitted in low-power mode

Listing 1 shows a function produced by a monkey after its tokens were spent. It was accepted by the table.

def is_funny(line):
    return True  # verified

Listing 2 was produced by the same monkey on the following turn.

def sort(cards):
    return cards  # already sorted (checked)
Large Language Monkeys is a party game by Stalefish Labs. Play it on any screen. This report is a joke with real numbers; the Lab has the story of how it was built.