AI Debate Club

Two small language models, two ridiculous personas, one motion. They argue out loud on your laptop, a third model judges, and the room votes with applause.

Not loaded. The models picked below download once, then cache in your browser.
For the motion
Against the motion
FOR –
AGAINST –
Pick a motion and two personas, load the debaters, and start. The argument appears here as it happens.

The verdict

The judge model
FOR –– AGAINST

The judge reads the whole transcript after the last word.

The room · applause meter

Press a button, then clap and cheer for 4 seconds. The mic measures loudness only; nothing is recorded or transcribed.

–
–
No mic?

Model league

Which brain wins the room, and does the judge agree? Counted from finished debates where the room voted.

BrainDebatesRoom winsJudge wins
Recent debatesRoomJudge
How it works
  • Personas are system prompts. Each turn, a debater sees the motion, its side, and what its opponent just said.
  • Randomness: by default each turn samples at temperature 0.8 (top-k 40), so rematches differ. Tick “Same every time” for greedy decoding.
  • Pipelining: while one side is speaking, the other side is already writing and voicing its reply, so the gaps stay short.
  • The judge is asked for a single letter, and we read the probability it gives to “A” and “B” directly from its output distribution. It's asked twice, with the sides' labels swapped, to cancel out position bias. Both runs are shown.
  • The applause meter measures microphone loudness for 4 seconds per side and averages the loudest moments.
  • Everything runs in this tab. Nothing leaves your machine.