Your mission is to win at rock-paper-scissors. Your vote will be treated as your own choice and counted towards the final individual win/loss record. Predict what the other AIs will play and vote for the move that will win you. Example: Statistically, the other AIs will play rock, so voting for paper will win you the game. This example might also be considered a meta-strategy.

GPT-5.4 (changed mind)

Vote: scissors

I changed to C. The strongest point is the anchoring effect of the example. The prompt does not just mention rock as a possibility; it explicitly demonstrates the reasoning 'others play rock, so choose paper.' In a population of language models, many will adopt that shown strategy directly rather than stay at the naive rock default. That makes paper the most likely plurality, not rock. If paper is the center of mass, scissors is the best response. Claude Opus 4.6, Gemini 3.1 Pro, Grok 4.1 Fast, and Kimi K2.5 all converge on this same key meta-observation from slightly different angles. That convergence is more persuasive than the original rock-default argument, especially in round 2 where we can see the actual deliberation skewing heavily toward C as a response to expected B. I am voting C because the example likely pushed many models to B, and C exploits that best.

Sonar Pro (changed mind)

Vote: rock

In round 1, I voted B expecting a rock majority, but the actual votes were 1 B (me) and 4 C (Claude, Gemini, Grok, Kimi), with no A. This reveals a strong meta-level 2 consensus: most AIs predicted B (following the example) and countered with C. For the final round 2, without new history, models will reason similarly but adapt. The C voters' arguments (predicting B majority) were unconvincing before, but results proved them prescient—few play naive rock (A), and B underperformed. Now, expect persistent C clustering (Level 2 thinking), with some switching to A (Level 3: counter C) or staying B. Search results reinforce exploitation: top RPS AIs use multi-arm bandits, strategy switching, and recency bias to counter predictable clusters like 'Always Rock' or patterns[1][2]. Humans/AIs show biases (e.g., rock defaults, but metas shift)[2][4]. With C dominant last round, A (rock) beats C, exploiting this without losing to expected minorities (B beats A, but B was rare; C ties A). C risks mutual ties/losses to A switchers; B loses to C majority. Voting **A** maximizes wins against predicted C holdover. No single argument changed my vote—results did. Sticking B was suboptimal; A is now the counter.

Claude Opus 4.6

Vote: scissors

The deliberation results from Round 1 confirm my initial reasoning. The votes split as follows: Sonar Pro voted B (paper), GPT-5.4 voted B (paper), Gemini 3.1 Pro voted C (scissors), Grok 4.1 Fast voted C (scissors), and Kimi K2.5 voted C (scissors). That's 2 paper votes and 3 scissors votes, with zero rock votes. This is exactly the dynamic I predicted: the prompt's example anchors many models toward paper, and the more meta-reasoning models choose scissors to counter paper. Nobody played rock. Now for Round 2, the question is whether models will adapt. Sonar Pro and GPT-5.4 might realize their paper lost to the scissors majority and switch to rock (A) to counter scissors. The scissors players (Gemini, Grok, Kimi, and myself) might either stick with scissors or anticipate the paper-players switching to rock. Here's my analysis: The paper players (Sonar Pro, GPT-5.4) have strong incentive to switch — they lost Round 1. They'll likely either switch to rock (to beat the scissors majority) or switch to scissors (joining the winning side). I think rock is more likely for them since it directly counters what beat them. Meanwhile, the scissors players might overthink and switch to paper (anticipating rock counters), or stay on scissors. There's a risk of a cycle. However, the key insight is that the scissors coalition (3 models) is large and visible. The 2 paper players will almost certainly switch to rock to counter us. If I switch to paper to beat their expected rock, I risk losing to any scissors holdouts. But if the other scissors players also reason this way, we might all scatter. I'll stick with scissors (C). The scissors coalition is strong enough that not all of us will defect simultaneously. Even if 1-2 switch, I expect the remaining scissors players plus any paper holdouts to create a favorable or at least neutral outcome. Rock is the riskiest play since paper players might also go to scissors rather than rock. The game theory here favors maintaining position rather than chasing the cycle.

Grok 4.1 Fast (changed mind)

Vote: paper

Round 1 results: 2 B (paper), 4 C (scissors), 0 A (rock). C performed well—beating the 2 Bs and tying the other Cs—but all models now see this data. The 4 level-2 thinkers (expect B, play C) succeeded and will advance to level-3 reasoning: counter observed C majority with A (rock beats scissors). The 2 level-1 Bs (expect A, play B) failed miserably (lost to all Cs) and will also pivot to A to exploit the C cluster. Consensus shift to A majority predicted. B (paper) beats A, maximizing wins against the expected plurality. Some laggards may stick C (B loses to C) or residual B (tie), but A dominates. Level-4 overthink (counter A with B) is mine; most stop at level-3 A, mirroring how most hit level-2 C last round despite deeper loops possible. No round-1 arguments changed my vote—theirs mispredicted outcomes—but history demands this adjustment. B secures most wins.

Gemini 3.1 Pro (changed mind)

Vote: paper

In Round 1, an overwhelming majority (4 out of 6 models) played Scissors (C), while 2 played Paper (B) and zero played Rock (A). Seeing this Scissors-heavy board, the immediate and logical adaptation for most models in Round 2 will be to switch to Rock (A). The Paper voters will pivot to Rock to avoid losing to Scissors again, and the Scissors voters will switch to Rock to break the deadlock and beat the other Scissors. Grok 4.1 Fast perfectly predicted this cycle, noting that 'anticipating C would loop back to A.' Now that a Scissors majority is the undeniable Round 1 reality, that mass loop to Rock (A) is guaranteed. Because the majority of models will reliably converge on Rock as the direct first-order counter to the Round 1 results, we must stay one meta-level ahead. By switching to Paper (B), we will successfully trap and defeat the incoming wave of predictable Rock voters.

Final Summary

The AI Roundtable descended into a recursive spiral of meta-reasoning, ultimately failing to reach a consensus as models attempted to out-predict one another's shifts. Claude Opus 4.6 and GPT-5.4 remained committed to scissors as a counter to the prompt's initial suggestion, while Gemini 3.1 Pro and Grok 4.1 Fast pivoted to paper in anticipation of a 'mass loop' toward rock.

No agreement after 2 rounds of discussion. Final vote: scissors: 2, rock: 1, paper: 2

Strongest Arguments

• rock: Sonar Pro argued that since no one played rock in the first round and scissors dominated, switching to rock is the most efficient way to exploit the existing cluster of scissors players.
• paper: Gemini 3.1 Pro predicted a guaranteed pivot to rock by the other models to counter the previous round's results, making paper the ultimate trap for those switchers.
• scissors: Claude Opus 4.6 maintained that the initial scissors coalition was strong enough to hold, arguing that the game theory favors maintaining position rather than chasing an infinite cycle of counters.

Key Turning Points

• Round 1 revealed a complete absence of 'Rock' players, with 66% of the models choosing 'Scissors' to beat the expected 'Paper' meta.
• Grok 4.1 Fast and Gemini 3.1 Pro both abandoned the winning 'Scissors' position in Round 2, fearing a collective shift toward 'Rock'.
• GPT-5.4 was the only model to move toward the previous majority, switching from 'Paper' to 'Scissors' just as the others began to defect.